| 1 | //===- SelectionDAG.cpp - Implement the SelectionDAG data structures ------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This implements the SelectionDAG class. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/CodeGen/SelectionDAG.h" |
| 14 | #include "SDNodeDbgValue.h" |
| 15 | #include "llvm/ADT/APFloat.h" |
| 16 | #include "llvm/ADT/APInt.h" |
| 17 | #include "llvm/ADT/APSInt.h" |
| 18 | #include "llvm/ADT/ArrayRef.h" |
| 19 | #include "llvm/ADT/BitVector.h" |
| 20 | #include "llvm/ADT/FoldingSet.h" |
| 21 | #include "llvm/ADT/None.h" |
| 22 | #include "llvm/ADT/STLExtras.h" |
| 23 | #include "llvm/ADT/SmallPtrSet.h" |
| 24 | #include "llvm/ADT/SmallVector.h" |
| 25 | #include "llvm/ADT/Triple.h" |
| 26 | #include "llvm/ADT/Twine.h" |
| 27 | #include "llvm/Analysis/ValueTracking.h" |
| 28 | #include "llvm/CodeGen/ISDOpcodes.h" |
| 29 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 30 | #include "llvm/CodeGen/MachineConstantPool.h" |
| 31 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 32 | #include "llvm/CodeGen/MachineFunction.h" |
| 33 | #include "llvm/CodeGen/MachineMemOperand.h" |
| 34 | #include "llvm/CodeGen/RuntimeLibcalls.h" |
| 35 | #include "llvm/CodeGen/SelectionDAGAddressAnalysis.h" |
| 36 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
| 37 | #include "llvm/CodeGen/SelectionDAGTargetInfo.h" |
| 38 | #include "llvm/CodeGen/TargetLowering.h" |
| 39 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 40 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 41 | #include "llvm/CodeGen/ValueTypes.h" |
| 42 | #include "llvm/IR/Constant.h" |
| 43 | #include "llvm/IR/Constants.h" |
| 44 | #include "llvm/IR/DataLayout.h" |
| 45 | #include "llvm/IR/DebugInfoMetadata.h" |
| 46 | #include "llvm/IR/DebugLoc.h" |
| 47 | #include "llvm/IR/DerivedTypes.h" |
| 48 | #include "llvm/IR/DiagnosticInfo.h" |
| 49 | #include "llvm/IR/Function.h" |
| 50 | #include "llvm/IR/GlobalValue.h" |
| 51 | #include "llvm/IR/Metadata.h" |
| 52 | #include "llvm/IR/Type.h" |
| 53 | #include "llvm/IR/Value.h" |
| 54 | #include "llvm/Support/Casting.h" |
| 55 | #include "llvm/Support/CodeGen.h" |
| 56 | #include "llvm/Support/Compiler.h" |
| 57 | #include "llvm/Support/Debug.h" |
| 58 | #include "llvm/Support/ErrorHandling.h" |
| 59 | #include "llvm/Support/KnownBits.h" |
| 60 | #include "llvm/Support/MachineValueType.h" |
| 61 | #include "llvm/Support/ManagedStatic.h" |
| 62 | #include "llvm/Support/MathExtras.h" |
| 63 | #include "llvm/Support/Mutex.h" |
| 64 | #include "llvm/Support/raw_ostream.h" |
| 65 | #include "llvm/Target/TargetMachine.h" |
| 66 | #include "llvm/Target/TargetOptions.h" |
| 67 | #include "llvm/Transforms/Utils/CheriSetBounds.h" |
| 68 | #include <algorithm> |
| 69 | #include <cassert> |
| 70 | #include <cstdint> |
| 71 | #include <cstdlib> |
| 72 | #include <limits> |
| 73 | #include <set> |
| 74 | #include <string> |
| 75 | #include <utility> |
| 76 | #include <vector> |
| 77 | |
| 78 | using namespace llvm; |
| 79 | |
| 80 | /// makeVTList - Return an instance of the SDVTList struct initialized with the |
| 81 | /// specified members. |
| 82 | static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) { |
| 83 | SDVTList Res = {VTs, NumVTs}; |
| 84 | return Res; |
| 85 | } |
| 86 | |
| 87 | // Default null implementations of the callbacks. |
| 88 | void SelectionDAG::DAGUpdateListener::NodeDeleted(SDNode*, SDNode*) {} |
| 89 | void SelectionDAG::DAGUpdateListener::NodeUpdated(SDNode*) {} |
| 90 | void SelectionDAG::DAGUpdateListener::NodeInserted(SDNode *) {} |
| 91 | |
| 92 | void SelectionDAG::DAGNodeDeletedListener::anchor() {} |
| 93 | |
| 94 | #define DEBUG_TYPE "selectiondag" |
| 95 | |
| 96 | static cl::opt<bool> EnableMemCpyDAGOpt("enable-memcpy-dag-opt" , |
| 97 | cl::Hidden, cl::init(true), |
| 98 | cl::desc("Gang up loads and stores generated by inlining of memcpy" )); |
| 99 | |
| 100 | static cl::opt<int> MaxLdStGlue("ldstmemcpy-glue-max" , |
| 101 | cl::desc("Number limit for gluing ld/st of memcpy." ), |
| 102 | cl::Hidden, cl::init(0)); |
| 103 | |
| 104 | static void NewSDValueDbgMsg(SDValue V, StringRef Msg, SelectionDAG *G) { |
| 105 | LLVM_DEBUG(dbgs() << Msg; V.getNode()->dump(G);); |
| 106 | } |
| 107 | |
| 108 | //===----------------------------------------------------------------------===// |
| 109 | // ConstantFPSDNode Class |
| 110 | //===----------------------------------------------------------------------===// |
| 111 | |
| 112 | /// isExactlyValue - We don't rely on operator== working on double values, as |
| 113 | /// it returns true for things that are clearly not equal, like -0.0 and 0.0. |
| 114 | /// As such, this method can be used to do an exact bit-for-bit comparison of |
| 115 | /// two floating point values. |
| 116 | bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const { |
| 117 | return getValueAPF().bitwiseIsEqual(V); |
| 118 | } |
| 119 | |
| 120 | bool ConstantFPSDNode::isValueValidForType(EVT VT, |
| 121 | const APFloat& Val) { |
| 122 | assert(VT.isFloatingPoint() && "Can only convert between FP types" ); |
| 123 | |
| 124 | // convert modifies in place, so make a copy. |
| 125 | APFloat Val2 = APFloat(Val); |
| 126 | bool losesInfo; |
| 127 | (void) Val2.convert(SelectionDAG::EVTToAPFloatSemantics(VT), |
| 128 | APFloat::rmNearestTiesToEven, |
| 129 | &losesInfo); |
| 130 | return !losesInfo; |
| 131 | } |
| 132 | |
| 133 | //===----------------------------------------------------------------------===// |
| 134 | // ISD Namespace |
| 135 | //===----------------------------------------------------------------------===// |
| 136 | |
| 137 | bool ISD::isConstantSplatVector(const SDNode *N, APInt &SplatVal) { |
| 138 | auto *BV = dyn_cast<BuildVectorSDNode>(N); |
| 139 | if (!BV) |
| 140 | return false; |
| 141 | |
| 142 | APInt SplatUndef; |
| 143 | unsigned SplatBitSize; |
| 144 | bool HasUndefs; |
| 145 | unsigned EltSize = N->getValueType(0).getVectorElementType().getSizeInBits(); |
| 146 | return BV->isConstantSplat(SplatVal, SplatUndef, SplatBitSize, HasUndefs, |
| 147 | EltSize) && |
| 148 | EltSize == SplatBitSize; |
| 149 | } |
| 150 | |
| 151 | // FIXME: AllOnes and AllZeros duplicate a lot of code. Could these be |
| 152 | // specializations of the more general isConstantSplatVector()? |
| 153 | |
| 154 | bool ISD::isBuildVectorAllOnes(const SDNode *N) { |
| 155 | // Look through a bit convert. |
| 156 | while (N->getOpcode() == ISD::BITCAST) |
| 157 | N = N->getOperand(0).getNode(); |
| 158 | |
| 159 | if (N->getOpcode() != ISD::BUILD_VECTOR) return false; |
| 160 | |
| 161 | unsigned i = 0, e = N->getNumOperands(); |
| 162 | |
| 163 | // Skip over all of the undef values. |
| 164 | while (i != e && N->getOperand(i).isUndef()) |
| 165 | ++i; |
| 166 | |
| 167 | // Do not accept an all-undef vector. |
| 168 | if (i == e) return false; |
| 169 | |
| 170 | // Do not accept build_vectors that aren't all constants or which have non-~0 |
| 171 | // elements. We have to be a bit careful here, as the type of the constant |
| 172 | // may not be the same as the type of the vector elements due to type |
| 173 | // legalization (the elements are promoted to a legal type for the target and |
| 174 | // a vector of a type may be legal when the base element type is not). |
| 175 | // We only want to check enough bits to cover the vector elements, because |
| 176 | // we care if the resultant vector is all ones, not whether the individual |
| 177 | // constants are. |
| 178 | SDValue NotZero = N->getOperand(i); |
| 179 | unsigned EltSize = N->getValueType(0).getScalarSizeInBits(); |
| 180 | if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(NotZero)) { |
| 181 | if (CN->getAPIntValue().countTrailingOnes() < EltSize) |
| 182 | return false; |
| 183 | } else if (ConstantFPSDNode *CFPN = dyn_cast<ConstantFPSDNode>(NotZero)) { |
| 184 | if (CFPN->getValueAPF().bitcastToAPInt().countTrailingOnes() < EltSize) |
| 185 | return false; |
| 186 | } else |
| 187 | return false; |
| 188 | |
| 189 | // Okay, we have at least one ~0 value, check to see if the rest match or are |
| 190 | // undefs. Even with the above element type twiddling, this should be OK, as |
| 191 | // the same type legalization should have applied to all the elements. |
| 192 | for (++i; i != e; ++i) |
| 193 | if (N->getOperand(i) != NotZero && !N->getOperand(i).isUndef()) |
| 194 | return false; |
| 195 | return true; |
| 196 | } |
| 197 | |
| 198 | bool ISD::isBuildVectorAllZeros(const SDNode *N) { |
| 199 | // Look through a bit convert. |
| 200 | while (N->getOpcode() == ISD::BITCAST) |
| 201 | N = N->getOperand(0).getNode(); |
| 202 | |
| 203 | if (N->getOpcode() != ISD::BUILD_VECTOR) return false; |
| 204 | |
| 205 | bool IsAllUndef = true; |
| 206 | for (const SDValue &Op : N->op_values()) { |
| 207 | if (Op.isUndef()) |
| 208 | continue; |
| 209 | IsAllUndef = false; |
| 210 | // Do not accept build_vectors that aren't all constants or which have non-0 |
| 211 | // elements. We have to be a bit careful here, as the type of the constant |
| 212 | // may not be the same as the type of the vector elements due to type |
| 213 | // legalization (the elements are promoted to a legal type for the target |
| 214 | // and a vector of a type may be legal when the base element type is not). |
| 215 | // We only want to check enough bits to cover the vector elements, because |
| 216 | // we care if the resultant vector is all zeros, not whether the individual |
| 217 | // constants are. |
| 218 | unsigned EltSize = N->getValueType(0).getScalarSizeInBits(); |
| 219 | if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op)) { |
| 220 | if (CN->getAPIntValue().countTrailingZeros() < EltSize) |
| 221 | return false; |
| 222 | } else if (ConstantFPSDNode *CFPN = dyn_cast<ConstantFPSDNode>(Op)) { |
| 223 | if (CFPN->getValueAPF().bitcastToAPInt().countTrailingZeros() < EltSize) |
| 224 | return false; |
| 225 | } else |
| 226 | return false; |
| 227 | } |
| 228 | |
| 229 | // Do not accept an all-undef vector. |
| 230 | if (IsAllUndef) |
| 231 | return false; |
| 232 | return true; |
| 233 | } |
| 234 | |
| 235 | bool ISD::isBuildVectorOfConstantSDNodes(const SDNode *N) { |
| 236 | if (N->getOpcode() != ISD::BUILD_VECTOR) |
| 237 | return false; |
| 238 | |
| 239 | for (const SDValue &Op : N->op_values()) { |
| 240 | if (Op.isUndef()) |
| 241 | continue; |
| 242 | if (!isa<ConstantSDNode>(Op)) |
| 243 | return false; |
| 244 | } |
| 245 | return true; |
| 246 | } |
| 247 | |
| 248 | bool ISD::isBuildVectorOfConstantFPSDNodes(const SDNode *N) { |
| 249 | if (N->getOpcode() != ISD::BUILD_VECTOR) |
| 250 | return false; |
| 251 | |
| 252 | for (const SDValue &Op : N->op_values()) { |
| 253 | if (Op.isUndef()) |
| 254 | continue; |
| 255 | if (!isa<ConstantFPSDNode>(Op)) |
| 256 | return false; |
| 257 | } |
| 258 | return true; |
| 259 | } |
| 260 | |
| 261 | bool ISD::allOperandsUndef(const SDNode *N) { |
| 262 | // Return false if the node has no operands. |
| 263 | // This is "logically inconsistent" with the definition of "all" but |
| 264 | // is probably the desired behavior. |
| 265 | if (N->getNumOperands() == 0) |
| 266 | return false; |
| 267 | return all_of(N->op_values(), [](SDValue Op) { return Op.isUndef(); }); |
| 268 | } |
| 269 | |
| 270 | bool ISD::matchUnaryPredicate(SDValue Op, |
| 271 | std::function<bool(ConstantSDNode *)> Match, |
| 272 | bool AllowUndefs) { |
| 273 | // FIXME: Add support for scalar UNDEF cases? |
| 274 | if (auto *Cst = dyn_cast<ConstantSDNode>(Op)) |
| 275 | return Match(Cst); |
| 276 | |
| 277 | // FIXME: Add support for vector UNDEF cases? |
| 278 | if (ISD::BUILD_VECTOR != Op.getOpcode()) |
| 279 | return false; |
| 280 | |
| 281 | EVT SVT = Op.getValueType().getScalarType(); |
| 282 | for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) { |
| 283 | if (AllowUndefs && Op.getOperand(i).isUndef()) { |
| 284 | if (!Match(nullptr)) |
| 285 | return false; |
| 286 | continue; |
| 287 | } |
| 288 | |
| 289 | auto *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(i)); |
| 290 | if (!Cst || Cst->getValueType(0) != SVT || !Match(Cst)) |
| 291 | return false; |
| 292 | } |
| 293 | return true; |
| 294 | } |
| 295 | |
| 296 | bool ISD::matchBinaryPredicate( |
| 297 | SDValue LHS, SDValue RHS, |
| 298 | std::function<bool(ConstantSDNode *, ConstantSDNode *)> Match, |
| 299 | bool AllowUndefs) { |
| 300 | if (LHS.getValueType() != RHS.getValueType()) |
| 301 | return false; |
| 302 | |
| 303 | // TODO: Add support for scalar UNDEF cases? |
| 304 | if (auto *LHSCst = dyn_cast<ConstantSDNode>(LHS)) |
| 305 | if (auto *RHSCst = dyn_cast<ConstantSDNode>(RHS)) |
| 306 | return Match(LHSCst, RHSCst); |
| 307 | |
| 308 | // TODO: Add support for vector UNDEF cases? |
| 309 | if (ISD::BUILD_VECTOR != LHS.getOpcode() || |
| 310 | ISD::BUILD_VECTOR != RHS.getOpcode()) |
| 311 | return false; |
| 312 | |
| 313 | EVT SVT = LHS.getValueType().getScalarType(); |
| 314 | for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) { |
| 315 | SDValue LHSOp = LHS.getOperand(i); |
| 316 | SDValue RHSOp = RHS.getOperand(i); |
| 317 | bool LHSUndef = AllowUndefs && LHSOp.isUndef(); |
| 318 | bool RHSUndef = AllowUndefs && RHSOp.isUndef(); |
| 319 | auto *LHSCst = dyn_cast<ConstantSDNode>(LHSOp); |
| 320 | auto *RHSCst = dyn_cast<ConstantSDNode>(RHSOp); |
| 321 | if ((!LHSCst && !LHSUndef) || (!RHSCst && !RHSUndef)) |
| 322 | return false; |
| 323 | if (LHSOp.getValueType() != SVT || |
| 324 | LHSOp.getValueType() != RHSOp.getValueType()) |
| 325 | return false; |
| 326 | if (!Match(LHSCst, RHSCst)) |
| 327 | return false; |
| 328 | } |
| 329 | return true; |
| 330 | } |
| 331 | |
| 332 | ISD::NodeType ISD::getExtForLoadExtType(bool IsFP, ISD::LoadExtType ExtType) { |
| 333 | switch (ExtType) { |
| 334 | case ISD::EXTLOAD: |
| 335 | return IsFP ? ISD::FP_EXTEND : ISD::ANY_EXTEND; |
| 336 | case ISD::SEXTLOAD: |
| 337 | return ISD::SIGN_EXTEND; |
| 338 | case ISD::ZEXTLOAD: |
| 339 | return ISD::ZERO_EXTEND; |
| 340 | default: |
| 341 | break; |
| 342 | } |
| 343 | |
| 344 | llvm_unreachable("Invalid LoadExtType" ); |
| 345 | } |
| 346 | |
| 347 | ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) { |
| 348 | // To perform this operation, we just need to swap the L and G bits of the |
| 349 | // operation. |
| 350 | unsigned OldL = (Operation >> 2) & 1; |
| 351 | unsigned OldG = (Operation >> 1) & 1; |
| 352 | return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits |
| 353 | (OldL << 1) | // New G bit |
| 354 | (OldG << 2)); // New L bit. |
| 355 | } |
| 356 | |
| 357 | ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, EVT VT) { |
| 358 | bool IsInteger = VT.isInteger(); |
| 359 | |
| 360 | unsigned Operation = Op; |
| 361 | if (IsInteger || VT.isFatPointer()) |
| 362 | Operation ^= 7; // Flip L, G, E bits, but not U. |
| 363 | else |
| 364 | Operation ^= 15; // Flip all of the condition bits. |
| 365 | |
| 366 | if (Operation > ISD::SETTRUE2) |
| 367 | Operation &= ~8; // Don't let N and U bits get set. |
| 368 | |
| 369 | return ISD::CondCode(Operation); |
| 370 | } |
| 371 | |
| 372 | /// For an integer comparison, return 1 if the comparison is a signed operation |
| 373 | /// and 2 if the result is an unsigned comparison. Return zero if the operation |
| 374 | /// does not depend on the sign of the input (setne and seteq). |
| 375 | static int isSignedOp(ISD::CondCode Opcode) { |
| 376 | switch (Opcode) { |
| 377 | default: llvm_unreachable("Illegal integer setcc operation!" ); |
| 378 | case ISD::SETEQ: |
| 379 | case ISD::SETNE: return 0; |
| 380 | case ISD::SETLT: |
| 381 | case ISD::SETLE: |
| 382 | case ISD::SETGT: |
| 383 | case ISD::SETGE: return 1; |
| 384 | case ISD::SETULT: |
| 385 | case ISD::SETULE: |
| 386 | case ISD::SETUGT: |
| 387 | case ISD::SETUGE: return 2; |
| 388 | } |
| 389 | } |
| 390 | |
| 391 | ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2, |
| 392 | EVT VT) { |
| 393 | // XXXAR: I don't think we can fold setcc or operations with fat pointers |
| 394 | if (VT.isFatPointer()) |
| 395 | return ISD::SETCC_INVALID; |
| 396 | bool IsInteger = VT.isInteger(); |
| 397 | if (IsInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3) |
| 398 | // Cannot fold a signed integer setcc with an unsigned integer setcc. |
| 399 | return ISD::SETCC_INVALID; |
| 400 | |
| 401 | unsigned Op = Op1 | Op2; // Combine all of the condition bits. |
| 402 | |
| 403 | // If the N and U bits get set, then the resultant comparison DOES suddenly |
| 404 | // care about orderedness, and it is true when ordered. |
| 405 | if (Op > ISD::SETTRUE2) |
| 406 | Op &= ~16; // Clear the U bit if the N bit is set. |
| 407 | |
| 408 | // Canonicalize illegal integer setcc's. |
| 409 | if (IsInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT |
| 410 | Op = ISD::SETNE; |
| 411 | |
| 412 | return ISD::CondCode(Op); |
| 413 | } |
| 414 | |
| 415 | ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2, |
| 416 | EVT VT) { |
| 417 | // XXXAR: I don't think we can fold setcc and operations with fat pointers |
| 418 | if (VT.isFatPointer()) |
| 419 | return ISD::SETCC_INVALID; |
| 420 | bool IsInteger = VT.isInteger(); |
| 421 | if (IsInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3) |
| 422 | // Cannot fold a signed setcc with an unsigned setcc. |
| 423 | return ISD::SETCC_INVALID; |
| 424 | |
| 425 | // Combine all of the condition bits. |
| 426 | ISD::CondCode Result = ISD::CondCode(Op1 & Op2); |
| 427 | |
| 428 | // Canonicalize illegal integer setcc's. |
| 429 | if (IsInteger) { |
| 430 | switch (Result) { |
| 431 | default: break; |
| 432 | case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT |
| 433 | case ISD::SETOEQ: // SETEQ & SETU[LG]E |
| 434 | case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE |
| 435 | case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE |
| 436 | case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | return Result; |
| 441 | } |
| 442 | |
| 443 | //===----------------------------------------------------------------------===// |
| 444 | // SDNode Profile Support |
| 445 | //===----------------------------------------------------------------------===// |
| 446 | |
| 447 | /// AddNodeIDOpcode - Add the node opcode to the NodeID data. |
| 448 | static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) { |
| 449 | ID.AddInteger(OpC); |
| 450 | } |
| 451 | |
| 452 | /// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them |
| 453 | /// solely with their pointer. |
| 454 | static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) { |
| 455 | ID.AddPointer(VTList.VTs); |
| 456 | } |
| 457 | |
| 458 | /// AddNodeIDOperands - Various routines for adding operands to the NodeID data. |
| 459 | static void AddNodeIDOperands(FoldingSetNodeID &ID, |
| 460 | ArrayRef<SDValue> Ops) { |
| 461 | for (auto& Op : Ops) { |
| 462 | ID.AddPointer(Op.getNode()); |
| 463 | ID.AddInteger(Op.getResNo()); |
| 464 | } |
| 465 | } |
| 466 | |
| 467 | /// AddNodeIDOperands - Various routines for adding operands to the NodeID data. |
| 468 | static void AddNodeIDOperands(FoldingSetNodeID &ID, |
| 469 | ArrayRef<SDUse> Ops) { |
| 470 | for (auto& Op : Ops) { |
| 471 | ID.AddPointer(Op.getNode()); |
| 472 | ID.AddInteger(Op.getResNo()); |
| 473 | } |
| 474 | } |
| 475 | |
| 476 | static void AddNodeIDNode(FoldingSetNodeID &ID, unsigned short OpC, |
| 477 | SDVTList VTList, ArrayRef<SDValue> OpList) { |
| 478 | AddNodeIDOpcode(ID, OpC); |
| 479 | AddNodeIDValueTypes(ID, VTList); |
| 480 | AddNodeIDOperands(ID, OpList); |
| 481 | } |
| 482 | |
| 483 | /// If this is an SDNode with special info, add this info to the NodeID data. |
| 484 | static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) { |
| 485 | switch (N->getOpcode()) { |
| 486 | case ISD::TargetExternalSymbol: |
| 487 | case ISD::ExternalSymbol: |
| 488 | case ISD::MCSymbol: |
| 489 | llvm_unreachable("Should only be used on nodes with operands" ); |
| 490 | default: break; // Normal nodes don't need extra info. |
| 491 | case ISD::TargetConstant: |
| 492 | case ISD::Constant: { |
| 493 | const ConstantSDNode *C = cast<ConstantSDNode>(N); |
| 494 | ID.AddPointer(C->getConstantIntValue()); |
| 495 | ID.AddBoolean(C->isOpaque()); |
| 496 | break; |
| 497 | } |
| 498 | case ISD::TargetConstantFP: |
| 499 | case ISD::ConstantFP: |
| 500 | ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue()); |
| 501 | break; |
| 502 | case ISD::TargetGlobalAddress: |
| 503 | case ISD::GlobalAddress: |
| 504 | case ISD::TargetGlobalTLSAddress: |
| 505 | case ISD::GlobalTLSAddress: { |
| 506 | const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N); |
| 507 | ID.AddPointer(GA->getGlobal()); |
| 508 | ID.AddInteger(GA->getOffset()); |
| 509 | ID.AddInteger(GA->getTargetFlags()); |
| 510 | break; |
| 511 | } |
| 512 | case ISD::BasicBlock: |
| 513 | ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock()); |
| 514 | break; |
| 515 | case ISD::Register: |
| 516 | ID.AddInteger(cast<RegisterSDNode>(N)->getReg()); |
| 517 | break; |
| 518 | case ISD::RegisterMask: |
| 519 | ID.AddPointer(cast<RegisterMaskSDNode>(N)->getRegMask()); |
| 520 | break; |
| 521 | case ISD::SRCVALUE: |
| 522 | ID.AddPointer(cast<SrcValueSDNode>(N)->getValue()); |
| 523 | break; |
| 524 | case ISD::FrameIndex: |
| 525 | case ISD::TargetFrameIndex: |
| 526 | ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex()); |
| 527 | break; |
| 528 | case ISD::LIFETIME_START: |
| 529 | case ISD::LIFETIME_END: |
| 530 | if (cast<LifetimeSDNode>(N)->hasOffset()) { |
| 531 | ID.AddInteger(cast<LifetimeSDNode>(N)->getSize()); |
| 532 | ID.AddInteger(cast<LifetimeSDNode>(N)->getOffset()); |
| 533 | } |
| 534 | break; |
| 535 | case ISD::JumpTable: |
| 536 | case ISD::TargetJumpTable: |
| 537 | ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex()); |
| 538 | ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags()); |
| 539 | break; |
| 540 | case ISD::ConstantPool: |
| 541 | case ISD::TargetConstantPool: { |
| 542 | const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N); |
| 543 | ID.AddInteger(CP->getAlignment()); |
| 544 | ID.AddInteger(CP->getOffset()); |
| 545 | if (CP->isMachineConstantPoolEntry()) |
| 546 | CP->getMachineCPVal()->addSelectionDAGCSEId(ID); |
| 547 | else |
| 548 | ID.AddPointer(CP->getConstVal()); |
| 549 | ID.AddInteger(CP->getTargetFlags()); |
| 550 | break; |
| 551 | } |
| 552 | case ISD::TargetIndex: { |
| 553 | const TargetIndexSDNode *TI = cast<TargetIndexSDNode>(N); |
| 554 | ID.AddInteger(TI->getIndex()); |
| 555 | ID.AddInteger(TI->getOffset()); |
| 556 | ID.AddInteger(TI->getTargetFlags()); |
| 557 | break; |
| 558 | } |
| 559 | case ISD::LOAD: { |
| 560 | const LoadSDNode *LD = cast<LoadSDNode>(N); |
| 561 | ID.AddInteger(LD->getMemoryVT().getRawBits()); |
| 562 | ID.AddInteger(LD->getRawSubclassData()); |
| 563 | ID.AddInteger(LD->getPointerInfo().getAddrSpace()); |
| 564 | break; |
| 565 | } |
| 566 | case ISD::STORE: { |
| 567 | const StoreSDNode *ST = cast<StoreSDNode>(N); |
| 568 | ID.AddInteger(ST->getMemoryVT().getRawBits()); |
| 569 | ID.AddInteger(ST->getRawSubclassData()); |
| 570 | ID.AddInteger(ST->getPointerInfo().getAddrSpace()); |
| 571 | break; |
| 572 | } |
| 573 | case ISD::MLOAD: { |
| 574 | const MaskedLoadSDNode *MLD = cast<MaskedLoadSDNode>(N); |
| 575 | ID.AddInteger(MLD->getMemoryVT().getRawBits()); |
| 576 | ID.AddInteger(MLD->getRawSubclassData()); |
| 577 | ID.AddInteger(MLD->getPointerInfo().getAddrSpace()); |
| 578 | break; |
| 579 | } |
| 580 | case ISD::MSTORE: { |
| 581 | const MaskedStoreSDNode *MST = cast<MaskedStoreSDNode>(N); |
| 582 | ID.AddInteger(MST->getMemoryVT().getRawBits()); |
| 583 | ID.AddInteger(MST->getRawSubclassData()); |
| 584 | ID.AddInteger(MST->getPointerInfo().getAddrSpace()); |
| 585 | break; |
| 586 | } |
| 587 | case ISD::MGATHER: { |
| 588 | const MaskedGatherSDNode *MG = cast<MaskedGatherSDNode>(N); |
| 589 | ID.AddInteger(MG->getMemoryVT().getRawBits()); |
| 590 | ID.AddInteger(MG->getRawSubclassData()); |
| 591 | ID.AddInteger(MG->getPointerInfo().getAddrSpace()); |
| 592 | break; |
| 593 | } |
| 594 | case ISD::MSCATTER: { |
| 595 | const MaskedScatterSDNode *MS = cast<MaskedScatterSDNode>(N); |
| 596 | ID.AddInteger(MS->getMemoryVT().getRawBits()); |
| 597 | ID.AddInteger(MS->getRawSubclassData()); |
| 598 | ID.AddInteger(MS->getPointerInfo().getAddrSpace()); |
| 599 | break; |
| 600 | } |
| 601 | case ISD::ATOMIC_CMP_SWAP: |
| 602 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: |
| 603 | case ISD::ATOMIC_SWAP: |
| 604 | case ISD::ATOMIC_LOAD_ADD: |
| 605 | case ISD::ATOMIC_LOAD_SUB: |
| 606 | case ISD::ATOMIC_LOAD_AND: |
| 607 | case ISD::ATOMIC_LOAD_CLR: |
| 608 | case ISD::ATOMIC_LOAD_OR: |
| 609 | case ISD::ATOMIC_LOAD_XOR: |
| 610 | case ISD::ATOMIC_LOAD_NAND: |
| 611 | case ISD::ATOMIC_LOAD_MIN: |
| 612 | case ISD::ATOMIC_LOAD_MAX: |
| 613 | case ISD::ATOMIC_LOAD_UMIN: |
| 614 | case ISD::ATOMIC_LOAD_UMAX: |
| 615 | case ISD::ATOMIC_LOAD: |
| 616 | case ISD::ATOMIC_STORE: { |
| 617 | const AtomicSDNode *AT = cast<AtomicSDNode>(N); |
| 618 | ID.AddInteger(AT->getMemoryVT().getRawBits()); |
| 619 | ID.AddInteger(AT->getRawSubclassData()); |
| 620 | ID.AddInteger(AT->getPointerInfo().getAddrSpace()); |
| 621 | break; |
| 622 | } |
| 623 | case ISD::PREFETCH: { |
| 624 | const MemSDNode *PF = cast<MemSDNode>(N); |
| 625 | ID.AddInteger(PF->getPointerInfo().getAddrSpace()); |
| 626 | break; |
| 627 | } |
| 628 | case ISD::VECTOR_SHUFFLE: { |
| 629 | const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); |
| 630 | for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements(); |
| 631 | i != e; ++i) |
| 632 | ID.AddInteger(SVN->getMaskElt(i)); |
| 633 | break; |
| 634 | } |
| 635 | case ISD::TargetBlockAddress: |
| 636 | case ISD::BlockAddress: { |
| 637 | const BlockAddressSDNode *BA = cast<BlockAddressSDNode>(N); |
| 638 | ID.AddPointer(BA->getBlockAddress()); |
| 639 | ID.AddInteger(BA->getOffset()); |
| 640 | ID.AddInteger(BA->getTargetFlags()); |
| 641 | break; |
| 642 | } |
| 643 | } // end switch (N->getOpcode()) |
| 644 | |
| 645 | // Target specific memory nodes could also have address spaces to check. |
| 646 | if (N->isTargetMemoryOpcode()) |
| 647 | ID.AddInteger(cast<MemSDNode>(N)->getPointerInfo().getAddrSpace()); |
| 648 | } |
| 649 | |
| 650 | /// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID |
| 651 | /// data. |
| 652 | static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) { |
| 653 | AddNodeIDOpcode(ID, N->getOpcode()); |
| 654 | // Add the return value info. |
| 655 | AddNodeIDValueTypes(ID, N->getVTList()); |
| 656 | // Add the operand info. |
| 657 | AddNodeIDOperands(ID, N->ops()); |
| 658 | |
| 659 | // Handle SDNode leafs with special info. |
| 660 | AddNodeIDCustom(ID, N); |
| 661 | } |
| 662 | |
| 663 | //===----------------------------------------------------------------------===// |
| 664 | // SelectionDAG Class |
| 665 | //===----------------------------------------------------------------------===// |
| 666 | |
| 667 | /// doNotCSE - Return true if CSE should not be performed for this node. |
| 668 | static bool doNotCSE(SDNode *N) { |
| 669 | if (N->getValueType(0) == MVT::Glue) |
| 670 | return true; // Never CSE anything that produces a flag. |
| 671 | |
| 672 | switch (N->getOpcode()) { |
| 673 | default: break; |
| 674 | case ISD::HANDLENODE: |
| 675 | case ISD::EH_LABEL: |
| 676 | return true; // Never CSE these nodes. |
| 677 | } |
| 678 | |
| 679 | // Check that remaining values produced are not flags. |
| 680 | for (unsigned i = 1, e = N->getNumValues(); i != e; ++i) |
| 681 | if (N->getValueType(i) == MVT::Glue) |
| 682 | return true; // Never CSE anything that produces a flag. |
| 683 | |
| 684 | return false; |
| 685 | } |
| 686 | |
| 687 | /// RemoveDeadNodes - This method deletes all unreachable nodes in the |
| 688 | /// SelectionDAG. |
| 689 | void SelectionDAG::RemoveDeadNodes() { |
| 690 | // Create a dummy node (which is not added to allnodes), that adds a reference |
| 691 | // to the root node, preventing it from being deleted. |
| 692 | HandleSDNode Dummy(getRoot()); |
| 693 | |
| 694 | SmallVector<SDNode*, 128> DeadNodes; |
| 695 | |
| 696 | // Add all obviously-dead nodes to the DeadNodes worklist. |
| 697 | for (SDNode &Node : allnodes()) |
| 698 | if (Node.use_empty()) |
| 699 | DeadNodes.push_back(&Node); |
| 700 | |
| 701 | RemoveDeadNodes(DeadNodes); |
| 702 | |
| 703 | // If the root changed (e.g. it was a dead load, update the root). |
| 704 | setRoot(Dummy.getValue()); |
| 705 | } |
| 706 | |
| 707 | /// RemoveDeadNodes - This method deletes the unreachable nodes in the |
| 708 | /// given list, and any nodes that become unreachable as a result. |
| 709 | void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes) { |
| 710 | |
| 711 | // Process the worklist, deleting the nodes and adding their uses to the |
| 712 | // worklist. |
| 713 | while (!DeadNodes.empty()) { |
| 714 | SDNode *N = DeadNodes.pop_back_val(); |
| 715 | // Skip to next node if we've already managed to delete the node. This could |
| 716 | // happen if replacing a node causes a node previously added to the node to |
| 717 | // be deleted. |
| 718 | if (N->getOpcode() == ISD::DELETED_NODE) |
| 719 | continue; |
| 720 | |
| 721 | for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next) |
| 722 | DUL->NodeDeleted(N, nullptr); |
| 723 | |
| 724 | // Take the node out of the appropriate CSE map. |
| 725 | RemoveNodeFromCSEMaps(N); |
| 726 | |
| 727 | // Next, brutally remove the operand list. This is safe to do, as there are |
| 728 | // no cycles in the graph. |
| 729 | for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) { |
| 730 | SDUse &Use = *I++; |
| 731 | SDNode *Operand = Use.getNode(); |
| 732 | Use.set(SDValue()); |
| 733 | |
| 734 | // Now that we removed this operand, see if there are no uses of it left. |
| 735 | if (Operand->use_empty()) |
| 736 | DeadNodes.push_back(Operand); |
| 737 | } |
| 738 | |
| 739 | DeallocateNode(N); |
| 740 | } |
| 741 | } |
| 742 | |
| 743 | void SelectionDAG::RemoveDeadNode(SDNode *N){ |
| 744 | SmallVector<SDNode*, 16> DeadNodes(1, N); |
| 745 | |
| 746 | // Create a dummy node that adds a reference to the root node, preventing |
| 747 | // it from being deleted. (This matters if the root is an operand of the |
| 748 | // dead node.) |
| 749 | HandleSDNode Dummy(getRoot()); |
| 750 | |
| 751 | RemoveDeadNodes(DeadNodes); |
| 752 | } |
| 753 | |
| 754 | void SelectionDAG::DeleteNode(SDNode *N) { |
| 755 | // First take this out of the appropriate CSE map. |
| 756 | RemoveNodeFromCSEMaps(N); |
| 757 | |
| 758 | // Finally, remove uses due to operands of this node, remove from the |
| 759 | // AllNodes list, and delete the node. |
| 760 | DeleteNodeNotInCSEMaps(N); |
| 761 | } |
| 762 | |
| 763 | void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) { |
| 764 | assert(N->getIterator() != AllNodes.begin() && |
| 765 | "Cannot delete the entry node!" ); |
| 766 | assert(N->use_empty() && "Cannot delete a node that is not dead!" ); |
| 767 | |
| 768 | // Drop all of the operands and decrement used node's use counts. |
| 769 | N->DropOperands(); |
| 770 | |
| 771 | DeallocateNode(N); |
| 772 | } |
| 773 | |
| 774 | void SDDbgInfo::erase(const SDNode *Node) { |
| 775 | DbgValMapType::iterator I = DbgValMap.find(Node); |
| 776 | if (I == DbgValMap.end()) |
| 777 | return; |
| 778 | for (auto &Val: I->second) |
| 779 | Val->setIsInvalidated(); |
| 780 | DbgValMap.erase(I); |
| 781 | } |
| 782 | |
| 783 | void SelectionDAG::DeallocateNode(SDNode *N) { |
| 784 | // If we have operands, deallocate them. |
| 785 | removeOperands(N); |
| 786 | |
| 787 | NodeAllocator.Deallocate(AllNodes.remove(N)); |
| 788 | |
| 789 | // Set the opcode to DELETED_NODE to help catch bugs when node |
| 790 | // memory is reallocated. |
| 791 | // FIXME: There are places in SDag that have grown a dependency on the opcode |
| 792 | // value in the released node. |
| 793 | __asan_unpoison_memory_region(&N->NodeType, sizeof(N->NodeType)); |
| 794 | N->NodeType = ISD::DELETED_NODE; |
| 795 | |
| 796 | // If any of the SDDbgValue nodes refer to this SDNode, invalidate |
| 797 | // them and forget about that node. |
| 798 | DbgInfo->erase(N); |
| 799 | } |
| 800 | |
| 801 | #ifndef NDEBUG |
| 802 | /// VerifySDNode - Sanity check the given SDNode. Aborts if it is invalid. |
| 803 | static void VerifySDNode(SDNode *N) { |
| 804 | switch (N->getOpcode()) { |
| 805 | default: |
| 806 | break; |
| 807 | case ISD::BUILD_PAIR: { |
| 808 | EVT VT = N->getValueType(0); |
| 809 | assert(N->getNumValues() == 1 && "Too many results!" ); |
| 810 | assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) && |
| 811 | "Wrong return type!" ); |
| 812 | assert(N->getNumOperands() == 2 && "Wrong number of operands!" ); |
| 813 | assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() && |
| 814 | "Mismatched operand types!" ); |
| 815 | assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() && |
| 816 | "Wrong operand type!" ); |
| 817 | assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() && |
| 818 | "Wrong return type size" ); |
| 819 | break; |
| 820 | } |
| 821 | case ISD::BUILD_VECTOR: { |
| 822 | assert(N->getNumValues() == 1 && "Too many results!" ); |
| 823 | assert(N->getValueType(0).isVector() && "Wrong return type!" ); |
| 824 | assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() && |
| 825 | "Wrong number of operands!" ); |
| 826 | EVT EltVT = N->getValueType(0).getVectorElementType(); |
| 827 | for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I) { |
| 828 | assert((I->getValueType() == EltVT || |
| 829 | (EltVT.isInteger() && I->getValueType().isInteger() && |
| 830 | EltVT.bitsLE(I->getValueType()))) && |
| 831 | "Wrong operand type!" ); |
| 832 | assert(I->getValueType() == N->getOperand(0).getValueType() && |
| 833 | "Operands must all have the same type" ); |
| 834 | } |
| 835 | break; |
| 836 | } |
| 837 | } |
| 838 | } |
| 839 | #endif // NDEBUG |
| 840 | |
| 841 | /// Insert a newly allocated node into the DAG. |
| 842 | /// |
| 843 | /// Handles insertion into the all nodes list and CSE map, as well as |
| 844 | /// verification and other common operations when a new node is allocated. |
| 845 | void SelectionDAG::InsertNode(SDNode *N) { |
| 846 | AllNodes.push_back(N); |
| 847 | #ifndef NDEBUG |
| 848 | N->PersistentId = NextPersistentId++; |
| 849 | VerifySDNode(N); |
| 850 | #endif |
| 851 | for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next) |
| 852 | DUL->NodeInserted(N); |
| 853 | } |
| 854 | |
| 855 | /// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that |
| 856 | /// correspond to it. This is useful when we're about to delete or repurpose |
| 857 | /// the node. We don't want future request for structurally identical nodes |
| 858 | /// to return N anymore. |
| 859 | bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) { |
| 860 | bool Erased = false; |
| 861 | switch (N->getOpcode()) { |
| 862 | case ISD::HANDLENODE: return false; // noop. |
| 863 | case ISD::CONDCODE: |
| 864 | assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] && |
| 865 | "Cond code doesn't exist!" ); |
| 866 | Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != nullptr; |
| 867 | CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = nullptr; |
| 868 | break; |
| 869 | case ISD::ExternalSymbol: |
| 870 | Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol()); |
| 871 | break; |
| 872 | case ISD::TargetExternalSymbol: { |
| 873 | ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N); |
| 874 | Erased = TargetExternalSymbols.erase( |
| 875 | std::pair<std::string,unsigned char>(ESN->getSymbol(), |
| 876 | ESN->getTargetFlags())); |
| 877 | break; |
| 878 | } |
| 879 | case ISD::MCSymbol: { |
| 880 | auto *MCSN = cast<MCSymbolSDNode>(N); |
| 881 | Erased = MCSymbols.erase(MCSN->getMCSymbol()); |
| 882 | break; |
| 883 | } |
| 884 | case ISD::VALUETYPE: { |
| 885 | EVT VT = cast<VTSDNode>(N)->getVT(); |
| 886 | if (VT.isExtended()) { |
| 887 | Erased = ExtendedValueTypeNodes.erase(VT); |
| 888 | } else { |
| 889 | Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != nullptr; |
| 890 | ValueTypeNodes[VT.getSimpleVT().SimpleTy] = nullptr; |
| 891 | } |
| 892 | break; |
| 893 | } |
| 894 | default: |
| 895 | // Remove it from the CSE Map. |
| 896 | assert(N->getOpcode() != ISD::DELETED_NODE && "DELETED_NODE in CSEMap!" ); |
| 897 | assert(N->getOpcode() != ISD::EntryToken && "EntryToken in CSEMap!" ); |
| 898 | Erased = CSEMap.RemoveNode(N); |
| 899 | break; |
| 900 | } |
| 901 | #ifndef NDEBUG |
| 902 | // Verify that the node was actually in one of the CSE maps, unless it has a |
| 903 | // flag result (which cannot be CSE'd) or is one of the special cases that are |
| 904 | // not subject to CSE. |
| 905 | if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Glue && |
| 906 | !N->isMachineOpcode() && !doNotCSE(N)) { |
| 907 | N->dump(this); |
| 908 | dbgs() << "\n" ; |
| 909 | llvm_unreachable("Node is not in map!" ); |
| 910 | } |
| 911 | #endif |
| 912 | return Erased; |
| 913 | } |
| 914 | |
| 915 | /// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE |
| 916 | /// maps and modified in place. Add it back to the CSE maps, unless an identical |
| 917 | /// node already exists, in which case transfer all its users to the existing |
| 918 | /// node. This transfer can potentially trigger recursive merging. |
| 919 | void |
| 920 | SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N) { |
| 921 | // For node types that aren't CSE'd, just act as if no identical node |
| 922 | // already exists. |
| 923 | if (!doNotCSE(N)) { |
| 924 | SDNode *Existing = CSEMap.GetOrInsertNode(N); |
| 925 | if (Existing != N) { |
| 926 | // If there was already an existing matching node, use ReplaceAllUsesWith |
| 927 | // to replace the dead one with the existing one. This can cause |
| 928 | // recursive merging of other unrelated nodes down the line. |
| 929 | ReplaceAllUsesWith(N, Existing); |
| 930 | |
| 931 | // N is now dead. Inform the listeners and delete it. |
| 932 | for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next) |
| 933 | DUL->NodeDeleted(N, Existing); |
| 934 | DeleteNodeNotInCSEMaps(N); |
| 935 | return; |
| 936 | } |
| 937 | } |
| 938 | |
| 939 | // If the node doesn't already exist, we updated it. Inform listeners. |
| 940 | for (DAGUpdateListener *DUL = UpdateListeners; DUL; DUL = DUL->Next) |
| 941 | DUL->NodeUpdated(N); |
| 942 | } |
| 943 | |
| 944 | /// FindModifiedNodeSlot - Find a slot for the specified node if its operands |
| 945 | /// were replaced with those specified. If this node is never memoized, |
| 946 | /// return null, otherwise return a pointer to the slot it would take. If a |
| 947 | /// node already exists with these operands, the slot will be non-null. |
| 948 | SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op, |
| 949 | void *&InsertPos) { |
| 950 | if (doNotCSE(N)) |
| 951 | return nullptr; |
| 952 | |
| 953 | SDValue Ops[] = { Op }; |
| 954 | FoldingSetNodeID ID; |
| 955 | AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops); |
| 956 | AddNodeIDCustom(ID, N); |
| 957 | SDNode *Node = FindNodeOrInsertPos(ID, SDLoc(N), InsertPos); |
| 958 | if (Node) |
| 959 | Node->intersectFlagsWith(N->getFlags()); |
| 960 | return Node; |
| 961 | } |
| 962 | |
| 963 | /// FindModifiedNodeSlot - Find a slot for the specified node if its operands |
| 964 | /// were replaced with those specified. If this node is never memoized, |
| 965 | /// return null, otherwise return a pointer to the slot it would take. If a |
| 966 | /// node already exists with these operands, the slot will be non-null. |
| 967 | SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, |
| 968 | SDValue Op1, SDValue Op2, |
| 969 | void *&InsertPos) { |
| 970 | if (doNotCSE(N)) |
| 971 | return nullptr; |
| 972 | |
| 973 | SDValue Ops[] = { Op1, Op2 }; |
| 974 | FoldingSetNodeID ID; |
| 975 | AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops); |
| 976 | AddNodeIDCustom(ID, N); |
| 977 | SDNode *Node = FindNodeOrInsertPos(ID, SDLoc(N), InsertPos); |
| 978 | if (Node) |
| 979 | Node->intersectFlagsWith(N->getFlags()); |
| 980 | return Node; |
| 981 | } |
| 982 | |
| 983 | /// FindModifiedNodeSlot - Find a slot for the specified node if its operands |
| 984 | /// were replaced with those specified. If this node is never memoized, |
| 985 | /// return null, otherwise return a pointer to the slot it would take. If a |
| 986 | /// node already exists with these operands, the slot will be non-null. |
| 987 | SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, ArrayRef<SDValue> Ops, |
| 988 | void *&InsertPos) { |
| 989 | if (doNotCSE(N)) |
| 990 | return nullptr; |
| 991 | |
| 992 | FoldingSetNodeID ID; |
| 993 | AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops); |
| 994 | AddNodeIDCustom(ID, N); |
| 995 | SDNode *Node = FindNodeOrInsertPos(ID, SDLoc(N), InsertPos); |
| 996 | if (Node) |
| 997 | Node->intersectFlagsWith(N->getFlags()); |
| 998 | return Node; |
| 999 | } |
| 1000 | |
| 1001 | unsigned SelectionDAG::getEVTAlignment(EVT VT) const { |
| 1002 | Type *Ty = VT == MVT::iPTR ? |
| 1003 | PointerType::get(Type::getInt8Ty(*getContext()), 0) : |
| 1004 | VT.getTypeForEVT(*getContext()); |
| 1005 | |
| 1006 | return getDataLayout().getABITypeAlignment(Ty); |
| 1007 | } |
| 1008 | |
| 1009 | // EntryNode could meaningfully have debug info if we can find it... |
| 1010 | SelectionDAG::SelectionDAG(const TargetMachine &tm, CodeGenOpt::Level OL) |
| 1011 | : TM(tm), OptLevel(OL), |
| 1012 | EntryNode(ISD::EntryToken, 0, DebugLoc(), getVTList(MVT::Other)), |
| 1013 | Root(getEntryNode()) { |
| 1014 | InsertNode(&EntryNode); |
| 1015 | DbgInfo = new SDDbgInfo(); |
| 1016 | } |
| 1017 | |
| 1018 | void SelectionDAG::(MachineFunction &NewMF, |
| 1019 | OptimizationRemarkEmitter &NewORE, |
| 1020 | Pass *PassPtr, const TargetLibraryInfo *LibraryInfo, |
| 1021 | LegacyDivergenceAnalysis * Divergence) { |
| 1022 | MF = &NewMF; |
| 1023 | SDAGISelPass = PassPtr; |
| 1024 | ORE = &NewORE; |
| 1025 | TLI = getSubtarget().getTargetLowering(); |
| 1026 | TSI = getSubtarget().getSelectionDAGInfo(); |
| 1027 | LibInfo = LibraryInfo; |
| 1028 | Context = &MF->getFunction().getContext(); |
| 1029 | DA = Divergence; |
| 1030 | } |
| 1031 | |
| 1032 | SelectionDAG::~SelectionDAG() { |
| 1033 | assert(!UpdateListeners && "Dangling registered DAGUpdateListeners" ); |
| 1034 | allnodes_clear(); |
| 1035 | OperandRecycler.clear(OperandAllocator); |
| 1036 | delete DbgInfo; |
| 1037 | } |
| 1038 | |
| 1039 | void SelectionDAG::allnodes_clear() { |
| 1040 | assert(&*AllNodes.begin() == &EntryNode); |
| 1041 | AllNodes.remove(AllNodes.begin()); |
| 1042 | while (!AllNodes.empty()) |
| 1043 | DeallocateNode(&AllNodes.front()); |
| 1044 | #ifndef NDEBUG |
| 1045 | NextPersistentId = 0; |
| 1046 | #endif |
| 1047 | } |
| 1048 | |
| 1049 | SDNode *SelectionDAG::FindNodeOrInsertPos(const FoldingSetNodeID &ID, |
| 1050 | void *&InsertPos) { |
| 1051 | SDNode *N = CSEMap.FindNodeOrInsertPos(ID, InsertPos); |
| 1052 | if (N) { |
| 1053 | switch (N->getOpcode()) { |
| 1054 | default: break; |
| 1055 | case ISD::Constant: |
| 1056 | case ISD::ConstantFP: |
| 1057 | llvm_unreachable("Querying for Constant and ConstantFP nodes requires " |
| 1058 | "debug location. Use another overload." ); |
| 1059 | } |
| 1060 | } |
| 1061 | return N; |
| 1062 | } |
| 1063 | |
| 1064 | SDNode *SelectionDAG::FindNodeOrInsertPos(const FoldingSetNodeID &ID, |
| 1065 | const SDLoc &DL, void *&InsertPos) { |
| 1066 | SDNode *N = CSEMap.FindNodeOrInsertPos(ID, InsertPos); |
| 1067 | if (N) { |
| 1068 | switch (N->getOpcode()) { |
| 1069 | case ISD::Constant: |
| 1070 | case ISD::ConstantFP: |
| 1071 | // Erase debug location from the node if the node is used at several |
| 1072 | // different places. Do not propagate one location to all uses as it |
| 1073 | // will cause a worse single stepping debugging experience. |
| 1074 | if (N->getDebugLoc() != DL.getDebugLoc()) |
| 1075 | N->setDebugLoc(DebugLoc()); |
| 1076 | break; |
| 1077 | default: |
| 1078 | // When the node's point of use is located earlier in the instruction |
| 1079 | // sequence than its prior point of use, update its debug info to the |
| 1080 | // earlier location. |
| 1081 | if (DL.getIROrder() && DL.getIROrder() < N->getIROrder()) |
| 1082 | N->setDebugLoc(DL.getDebugLoc()); |
| 1083 | break; |
| 1084 | } |
| 1085 | } |
| 1086 | return N; |
| 1087 | } |
| 1088 | |
| 1089 | void SelectionDAG::clear() { |
| 1090 | allnodes_clear(); |
| 1091 | OperandRecycler.clear(OperandAllocator); |
| 1092 | OperandAllocator.Reset(); |
| 1093 | CSEMap.clear(); |
| 1094 | |
| 1095 | ExtendedValueTypeNodes.clear(); |
| 1096 | ExternalSymbols.clear(); |
| 1097 | TargetExternalSymbols.clear(); |
| 1098 | MCSymbols.clear(); |
| 1099 | std::fill(CondCodeNodes.begin(), CondCodeNodes.end(), |
| 1100 | static_cast<CondCodeSDNode*>(nullptr)); |
| 1101 | std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(), |
| 1102 | static_cast<SDNode*>(nullptr)); |
| 1103 | |
| 1104 | EntryNode.UseList = nullptr; |
| 1105 | InsertNode(&EntryNode); |
| 1106 | Root = getEntryNode(); |
| 1107 | DbgInfo->clear(); |
| 1108 | } |
| 1109 | |
| 1110 | SDValue SelectionDAG::getFPExtendOrRound(SDValue Op, const SDLoc &DL, EVT VT) { |
| 1111 | return VT.bitsGT(Op.getValueType()) |
| 1112 | ? getNode(ISD::FP_EXTEND, DL, VT, Op) |
| 1113 | : getNode(ISD::FP_ROUND, DL, VT, Op, getIntPtrConstant(0, DL)); |
| 1114 | } |
| 1115 | |
| 1116 | SDValue SelectionDAG::getAnyExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) { |
| 1117 | return VT.bitsGT(Op.getValueType()) ? |
| 1118 | getNode(ISD::ANY_EXTEND, DL, VT, Op) : |
| 1119 | getNode(ISD::TRUNCATE, DL, VT, Op); |
| 1120 | } |
| 1121 | |
| 1122 | SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) { |
| 1123 | return VT.bitsGT(Op.getValueType()) ? |
| 1124 | getNode(ISD::SIGN_EXTEND, DL, VT, Op) : |
| 1125 | getNode(ISD::TRUNCATE, DL, VT, Op); |
| 1126 | } |
| 1127 | |
| 1128 | SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) { |
| 1129 | return VT.bitsGT(Op.getValueType()) ? |
| 1130 | getNode(ISD::ZERO_EXTEND, DL, VT, Op) : |
| 1131 | getNode(ISD::TRUNCATE, DL, VT, Op); |
| 1132 | } |
| 1133 | |
| 1134 | SDValue SelectionDAG::getBoolExtOrTrunc(SDValue Op, const SDLoc &SL, EVT VT, |
| 1135 | EVT OpVT) { |
| 1136 | if (VT.bitsLE(Op.getValueType())) |
| 1137 | return getNode(ISD::TRUNCATE, SL, VT, Op); |
| 1138 | |
| 1139 | TargetLowering::BooleanContent BType = TLI->getBooleanContents(OpVT); |
| 1140 | return getNode(TLI->getExtendForContent(BType), SL, VT, Op); |
| 1141 | } |
| 1142 | |
| 1143 | SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, const SDLoc &DL, EVT VT) { |
| 1144 | assert(!VT.isVector() && |
| 1145 | "getZeroExtendInReg should use the vector element type instead of " |
| 1146 | "the vector type!" ); |
| 1147 | if (Op.getValueType().getScalarType() == VT) return Op; |
| 1148 | unsigned BitWidth = Op.getScalarValueSizeInBits(); |
| 1149 | APInt Imm = APInt::getLowBitsSet(BitWidth, |
| 1150 | VT.getSizeInBits()); |
| 1151 | return getNode(ISD::AND, DL, Op.getValueType(), Op, |
| 1152 | getConstant(Imm, DL, Op.getValueType())); |
| 1153 | } |
| 1154 | |
| 1155 | SDValue SelectionDAG::getPtrExtOrTrunc(SDValue Op, const SDLoc &DL, EVT VT) { |
| 1156 | // Only unsigned pointer semantics are supported right now. In the future this |
| 1157 | // might delegate to TLI to check pointer signedness. |
| 1158 | return getZExtOrTrunc(Op, DL, VT); |
| 1159 | } |
| 1160 | |
| 1161 | SDValue SelectionDAG::getPtrExtendInReg(SDValue Op, const SDLoc &DL, EVT VT) { |
| 1162 | // Only unsigned pointer semantics are supported right now. In the future this |
| 1163 | // might delegate to TLI to check pointer signedness. |
| 1164 | return getZeroExtendInReg(Op, DL, VT); |
| 1165 | } |
| 1166 | |
| 1167 | /// getNOT - Create a bitwise NOT operation as (XOR Val, -1). |
| 1168 | SDValue SelectionDAG::getNOT(const SDLoc &DL, SDValue Val, EVT VT) { |
| 1169 | EVT EltVT = VT.getScalarType(); |
| 1170 | SDValue NegOne = |
| 1171 | getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), DL, VT); |
| 1172 | return getNode(ISD::XOR, DL, VT, Val, NegOne); |
| 1173 | } |
| 1174 | |
| 1175 | SDValue SelectionDAG::getLogicalNOT(const SDLoc &DL, SDValue Val, EVT VT) { |
| 1176 | SDValue TrueValue = getBoolConstant(true, DL, VT, VT); |
| 1177 | return getNode(ISD::XOR, DL, VT, Val, TrueValue); |
| 1178 | } |
| 1179 | |
| 1180 | SDValue SelectionDAG::getBoolConstant(bool V, const SDLoc &DL, EVT VT, |
| 1181 | EVT OpVT) { |
| 1182 | if (!V) |
| 1183 | return getConstant(0, DL, VT); |
| 1184 | |
| 1185 | switch (TLI->getBooleanContents(OpVT)) { |
| 1186 | case TargetLowering::ZeroOrOneBooleanContent: |
| 1187 | case TargetLowering::UndefinedBooleanContent: |
| 1188 | return getConstant(1, DL, VT); |
| 1189 | case TargetLowering::ZeroOrNegativeOneBooleanContent: |
| 1190 | return getAllOnesConstant(DL, VT); |
| 1191 | } |
| 1192 | llvm_unreachable("Unexpected boolean content enum!" ); |
| 1193 | } |
| 1194 | |
| 1195 | SDValue SelectionDAG::getNullCapability(const SDLoc &DL, EVT CapType) { |
| 1196 | assert(CapType.isFatPointer()); |
| 1197 | MVT IntVT = MVT::getIntegerVT(getDataLayout().getPointerSizeInBits(0)); |
| 1198 | return getNode(ISD::INTTOPTR, DL, CapType, getConstant(0, DL, IntVT)); |
| 1199 | } |
| 1200 | |
| 1201 | SDValue SelectionDAG::getConstant(uint64_t Val, const SDLoc &DL, EVT VT, |
| 1202 | bool isT, bool isO) { |
| 1203 | EVT EltVT = VT.getScalarType(); |
| 1204 | assert((EltVT.getSizeInBits() >= 64 || |
| 1205 | (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) && |
| 1206 | "getConstant with a uint64_t value that doesn't fit in the type!" ); |
| 1207 | return getConstant(APInt(EltVT.getSizeInBits(), Val), DL, VT, isT, isO); |
| 1208 | } |
| 1209 | |
| 1210 | SDValue SelectionDAG::getConstant(const APInt &Val, const SDLoc &DL, EVT VT, |
| 1211 | bool isT, bool isO) { |
| 1212 | return getConstant(*ConstantInt::get(*Context, Val), DL, VT, isT, isO); |
| 1213 | } |
| 1214 | |
| 1215 | SDValue SelectionDAG::getConstant(const ConstantInt &Val, const SDLoc &DL, |
| 1216 | EVT VT, bool isT, bool isO) { |
| 1217 | if (VT.isFatPointer()) { |
| 1218 | unsigned BitWidth = getDataLayout().getPointerSizeInBits(0); |
| 1219 | MVT IntVT = MVT::getIntegerVT(BitWidth); |
| 1220 | APInt Int = Val.getValue(); |
| 1221 | if (Int.getBitWidth() > BitWidth) |
| 1222 | Int = Int.trunc(BitWidth); |
| 1223 | const ConstantInt *V = ConstantInt::get(*Context, Int); |
| 1224 | SDValue IntVal = getConstant(*V, DL, IntVT, isT); |
| 1225 | assert(!isT && "Cannot create INTTOPTR targetconstant" ); |
| 1226 | // TODO: For MIPS we could copy from CNULL for value 0 |
| 1227 | return getNode(ISD::INTTOPTR, SDLoc(), VT, IntVal); |
| 1228 | } |
| 1229 | assert(VT.isInteger() && "Cannot create FP integer constant!" ); |
| 1230 | |
| 1231 | EVT EltVT = VT.getScalarType(); |
| 1232 | const ConstantInt *Elt = &Val; |
| 1233 | |
| 1234 | // In some cases the vector type is legal but the element type is illegal and |
| 1235 | // needs to be promoted, for example v8i8 on ARM. In this case, promote the |
| 1236 | // inserted value (the type does not need to match the vector element type). |
| 1237 | // Any extra bits introduced will be truncated away. |
| 1238 | if (VT.isVector() && TLI->getTypeAction(*getContext(), EltVT) == |
| 1239 | TargetLowering::TypePromoteInteger) { |
| 1240 | EltVT = TLI->getTypeToTransformTo(*getContext(), EltVT); |
| 1241 | APInt NewVal = Elt->getValue().zextOrTrunc(EltVT.getSizeInBits()); |
| 1242 | Elt = ConstantInt::get(*getContext(), NewVal); |
| 1243 | } |
| 1244 | // In other cases the element type is illegal and needs to be expanded, for |
| 1245 | // example v2i64 on MIPS32. In this case, find the nearest legal type, split |
| 1246 | // the value into n parts and use a vector type with n-times the elements. |
| 1247 | // Then bitcast to the type requested. |
| 1248 | // Legalizing constants too early makes the DAGCombiner's job harder so we |
| 1249 | // only legalize if the DAG tells us we must produce legal types. |
| 1250 | else if (NewNodesMustHaveLegalTypes && VT.isVector() && |
| 1251 | TLI->getTypeAction(*getContext(), EltVT) == |
| 1252 | TargetLowering::TypeExpandInteger) { |
| 1253 | const APInt &NewVal = Elt->getValue(); |
| 1254 | EVT ViaEltVT = TLI->getTypeToTransformTo(*getContext(), EltVT); |
| 1255 | unsigned ViaEltSizeInBits = ViaEltVT.getSizeInBits(); |
| 1256 | unsigned ViaVecNumElts = VT.getSizeInBits() / ViaEltSizeInBits; |
| 1257 | EVT ViaVecVT = EVT::getVectorVT(*getContext(), ViaEltVT, ViaVecNumElts); |
| 1258 | |
| 1259 | // Check the temporary vector is the correct size. If this fails then |
| 1260 | // getTypeToTransformTo() probably returned a type whose size (in bits) |
| 1261 | // isn't a power-of-2 factor of the requested type size. |
| 1262 | assert(ViaVecVT.getSizeInBits() == VT.getSizeInBits()); |
| 1263 | |
| 1264 | SmallVector<SDValue, 2> EltParts; |
| 1265 | for (unsigned i = 0; i < ViaVecNumElts / VT.getVectorNumElements(); ++i) { |
| 1266 | EltParts.push_back(getConstant(NewVal.lshr(i * ViaEltSizeInBits) |
| 1267 | .zextOrTrunc(ViaEltSizeInBits), DL, |
| 1268 | ViaEltVT, isT, isO)); |
| 1269 | } |
| 1270 | |
| 1271 | // EltParts is currently in little endian order. If we actually want |
| 1272 | // big-endian order then reverse it now. |
| 1273 | if (getDataLayout().isBigEndian()) |
| 1274 | std::reverse(EltParts.begin(), EltParts.end()); |
| 1275 | |
| 1276 | // The elements must be reversed when the element order is different |
| 1277 | // to the endianness of the elements (because the BITCAST is itself a |
| 1278 | // vector shuffle in this situation). However, we do not need any code to |
| 1279 | // perform this reversal because getConstant() is producing a vector |
| 1280 | // splat. |
| 1281 | // This situation occurs in MIPS MSA. |
| 1282 | |
| 1283 | SmallVector<SDValue, 8> Ops; |
| 1284 | for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) |
| 1285 | Ops.insert(Ops.end(), EltParts.begin(), EltParts.end()); |
| 1286 | |
| 1287 | SDValue V = getNode(ISD::BITCAST, DL, VT, getBuildVector(ViaVecVT, DL, Ops)); |
| 1288 | return V; |
| 1289 | } |
| 1290 | |
| 1291 | assert(Elt->getBitWidth() == EltVT.getSizeInBits() && |
| 1292 | "APInt size does not match type size!" ); |
| 1293 | unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant; |
| 1294 | FoldingSetNodeID ID; |
| 1295 | AddNodeIDNode(ID, Opc, getVTList(EltVT), None); |
| 1296 | ID.AddPointer(Elt); |
| 1297 | ID.AddBoolean(isO); |
| 1298 | void *IP = nullptr; |
| 1299 | SDNode *N = nullptr; |
| 1300 | if ((N = FindNodeOrInsertPos(ID, DL, IP))) |
| 1301 | if (!VT.isVector()) |
| 1302 | return SDValue(N, 0); |
| 1303 | |
| 1304 | if (!N) { |
| 1305 | N = newSDNode<ConstantSDNode>(isT, isO, Elt, EltVT); |
| 1306 | CSEMap.InsertNode(N, IP); |
| 1307 | InsertNode(N); |
| 1308 | NewSDValueDbgMsg(SDValue(N, 0), "Creating constant: " , this); |
| 1309 | } |
| 1310 | |
| 1311 | SDValue Result(N, 0); |
| 1312 | if (VT.isVector()) |
| 1313 | Result = getSplatBuildVector(VT, DL, Result); |
| 1314 | |
| 1315 | return Result; |
| 1316 | } |
| 1317 | |
| 1318 | SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, const SDLoc &DL, |
| 1319 | bool isTarget) { |
| 1320 | return getConstant(Val, DL, TLI->getPointerRangeTy(getDataLayout()), |
| 1321 | isTarget); |
| 1322 | } |
| 1323 | |
| 1324 | SDValue SelectionDAG::getShiftAmountConstant(uint64_t Val, EVT VT, |
| 1325 | const SDLoc &DL, bool LegalTypes) { |
| 1326 | EVT ShiftVT = TLI->getShiftAmountTy(VT, getDataLayout(), LegalTypes); |
| 1327 | return getConstant(Val, DL, ShiftVT); |
| 1328 | } |
| 1329 | |
| 1330 | SDValue SelectionDAG::getConstantFP(const APFloat &V, const SDLoc &DL, EVT VT, |
| 1331 | bool isTarget) { |
| 1332 | return getConstantFP(*ConstantFP::get(*getContext(), V), DL, VT, isTarget); |
| 1333 | } |
| 1334 | |
| 1335 | SDValue SelectionDAG::getConstantFP(const ConstantFP &V, const SDLoc &DL, |
| 1336 | EVT VT, bool isTarget) { |
| 1337 | assert(VT.isFloatingPoint() && "Cannot create integer FP constant!" ); |
| 1338 | |
| 1339 | EVT EltVT = VT.getScalarType(); |
| 1340 | |
| 1341 | // Do the map lookup using the actual bit pattern for the floating point |
| 1342 | // value, so that we don't have problems with 0.0 comparing equal to -0.0, and |
| 1343 | // we don't have issues with SNANs. |
| 1344 | unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP; |
| 1345 | FoldingSetNodeID ID; |
| 1346 | AddNodeIDNode(ID, Opc, getVTList(EltVT), None); |
| 1347 | ID.AddPointer(&V); |
| 1348 | void *IP = nullptr; |
| 1349 | SDNode *N = nullptr; |
| 1350 | if ((N = FindNodeOrInsertPos(ID, DL, IP))) |
| 1351 | if (!VT.isVector()) |
| 1352 | return SDValue(N, 0); |
| 1353 | |
| 1354 | if (!N) { |
| 1355 | N = newSDNode<ConstantFPSDNode>(isTarget, &V, EltVT); |
| 1356 | CSEMap.InsertNode(N, IP); |
| 1357 | InsertNode(N); |
| 1358 | } |
| 1359 | |
| 1360 | SDValue Result(N, 0); |
| 1361 | if (VT.isVector()) |
| 1362 | Result = getSplatBuildVector(VT, DL, Result); |
| 1363 | NewSDValueDbgMsg(Result, "Creating fp constant: " , this); |
| 1364 | return Result; |
| 1365 | } |
| 1366 | |
| 1367 | SDValue SelectionDAG::getConstantFP(double Val, const SDLoc &DL, EVT VT, |
| 1368 | bool isTarget) { |
| 1369 | EVT EltVT = VT.getScalarType(); |
| 1370 | if (EltVT == MVT::f32) |
| 1371 | return getConstantFP(APFloat((float)Val), DL, VT, isTarget); |
| 1372 | else if (EltVT == MVT::f64) |
| 1373 | return getConstantFP(APFloat(Val), DL, VT, isTarget); |
| 1374 | else if (EltVT == MVT::f80 || EltVT == MVT::f128 || EltVT == MVT::ppcf128 || |
| 1375 | EltVT == MVT::f16) { |
| 1376 | bool Ignored; |
| 1377 | APFloat APF = APFloat(Val); |
| 1378 | APF.convert(EVTToAPFloatSemantics(EltVT), APFloat::rmNearestTiesToEven, |
| 1379 | &Ignored); |
| 1380 | return getConstantFP(APF, DL, VT, isTarget); |
| 1381 | } else |
| 1382 | llvm_unreachable("Unsupported type in getConstantFP" ); |
| 1383 | } |
| 1384 | |
| 1385 | SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV, const SDLoc &DL, |
| 1386 | EVT VT, int64_t Offset, bool isTargetGA, |
| 1387 | unsigned char TargetFlags) { |
| 1388 | assert((TargetFlags == 0 || isTargetGA) && |
| 1389 | "Cannot set target flags on target-independent globals" ); |
| 1390 | |
| 1391 | // Truncate (with sign-extension) the offset value to the pointer size. |
| 1392 | unsigned BitWidth = getDataLayout().getPointerTypeSizeInBits(GV->getType()); |
| 1393 | if (BitWidth < 64) |
| 1394 | Offset = SignExtend64(Offset, BitWidth); |
| 1395 | |
| 1396 | unsigned Opc; |
| 1397 | if (GV->isThreadLocal()) |
| 1398 | Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress; |
| 1399 | else |
| 1400 | Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress; |
| 1401 | |
| 1402 | FoldingSetNodeID ID; |
| 1403 | AddNodeIDNode(ID, Opc, getVTList(VT), None); |
| 1404 | ID.AddPointer(GV); |
| 1405 | ID.AddInteger(Offset); |
| 1406 | ID.AddInteger(TargetFlags); |
| 1407 | void *IP = nullptr; |
| 1408 | if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) |
| 1409 | return SDValue(E, 0); |
| 1410 | |
| 1411 | auto *N = newSDNode<GlobalAddressSDNode>( |
| 1412 | Opc, DL.getIROrder(), DL.getDebugLoc(), GV, VT, Offset, TargetFlags); |
| 1413 | CSEMap.InsertNode(N, IP); |
| 1414 | InsertNode(N); |
| 1415 | return SDValue(N, 0); |
| 1416 | } |
| 1417 | |
| 1418 | SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) { |
| 1419 | unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex; |
| 1420 | FoldingSetNodeID ID; |
| 1421 | AddNodeIDNode(ID, Opc, getVTList(VT), None); |
| 1422 | ID.AddInteger(FI); |
| 1423 | void *IP = nullptr; |
| 1424 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1425 | return SDValue(E, 0); |
| 1426 | |
| 1427 | auto *N = newSDNode<FrameIndexSDNode>(FI, VT, isTarget); |
| 1428 | CSEMap.InsertNode(N, IP); |
| 1429 | InsertNode(N); |
| 1430 | return SDValue(N, 0); |
| 1431 | } |
| 1432 | |
| 1433 | SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget, |
| 1434 | unsigned char TargetFlags) { |
| 1435 | assert((TargetFlags == 0 || isTarget) && |
| 1436 | "Cannot set target flags on target-independent jump tables" ); |
| 1437 | unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable; |
| 1438 | FoldingSetNodeID ID; |
| 1439 | AddNodeIDNode(ID, Opc, getVTList(VT), None); |
| 1440 | ID.AddInteger(JTI); |
| 1441 | ID.AddInteger(TargetFlags); |
| 1442 | void *IP = nullptr; |
| 1443 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1444 | return SDValue(E, 0); |
| 1445 | |
| 1446 | auto *N = newSDNode<JumpTableSDNode>(JTI, VT, isTarget, TargetFlags); |
| 1447 | CSEMap.InsertNode(N, IP); |
| 1448 | InsertNode(N); |
| 1449 | return SDValue(N, 0); |
| 1450 | } |
| 1451 | |
| 1452 | SDValue SelectionDAG::getConstantPool(const Constant *C, EVT VT, |
| 1453 | unsigned Alignment, int Offset, |
| 1454 | bool isTarget, |
| 1455 | unsigned char TargetFlags) { |
| 1456 | assert((TargetFlags == 0 || isTarget) && |
| 1457 | "Cannot set target flags on target-independent globals" ); |
| 1458 | if (Alignment == 0) |
| 1459 | Alignment = MF->getFunction().hasOptSize() |
| 1460 | ? getDataLayout().getABITypeAlignment(C->getType()) |
| 1461 | : getDataLayout().getPrefTypeAlignment(C->getType()); |
| 1462 | unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool; |
| 1463 | FoldingSetNodeID ID; |
| 1464 | AddNodeIDNode(ID, Opc, getVTList(VT), None); |
| 1465 | ID.AddInteger(Alignment); |
| 1466 | ID.AddInteger(Offset); |
| 1467 | ID.AddPointer(C); |
| 1468 | ID.AddInteger(TargetFlags); |
| 1469 | void *IP = nullptr; |
| 1470 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1471 | return SDValue(E, 0); |
| 1472 | |
| 1473 | auto *N = newSDNode<ConstantPoolSDNode>(isTarget, C, VT, Offset, Alignment, |
| 1474 | TargetFlags); |
| 1475 | CSEMap.InsertNode(N, IP); |
| 1476 | InsertNode(N); |
| 1477 | return SDValue(N, 0); |
| 1478 | } |
| 1479 | |
| 1480 | SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT, |
| 1481 | unsigned Alignment, int Offset, |
| 1482 | bool isTarget, |
| 1483 | unsigned char TargetFlags) { |
| 1484 | assert((TargetFlags == 0 || isTarget) && |
| 1485 | "Cannot set target flags on target-independent globals" ); |
| 1486 | if (Alignment == 0) |
| 1487 | Alignment = getDataLayout().getPrefTypeAlignment(C->getType()); |
| 1488 | unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool; |
| 1489 | FoldingSetNodeID ID; |
| 1490 | AddNodeIDNode(ID, Opc, getVTList(VT), None); |
| 1491 | ID.AddInteger(Alignment); |
| 1492 | ID.AddInteger(Offset); |
| 1493 | C->addSelectionDAGCSEId(ID); |
| 1494 | ID.AddInteger(TargetFlags); |
| 1495 | void *IP = nullptr; |
| 1496 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1497 | return SDValue(E, 0); |
| 1498 | |
| 1499 | auto *N = newSDNode<ConstantPoolSDNode>(isTarget, C, VT, Offset, Alignment, |
| 1500 | TargetFlags); |
| 1501 | CSEMap.InsertNode(N, IP); |
| 1502 | InsertNode(N); |
| 1503 | return SDValue(N, 0); |
| 1504 | } |
| 1505 | |
| 1506 | SDValue SelectionDAG::getTargetIndex(int Index, EVT VT, int64_t Offset, |
| 1507 | unsigned char TargetFlags) { |
| 1508 | FoldingSetNodeID ID; |
| 1509 | AddNodeIDNode(ID, ISD::TargetIndex, getVTList(VT), None); |
| 1510 | ID.AddInteger(Index); |
| 1511 | ID.AddInteger(Offset); |
| 1512 | ID.AddInteger(TargetFlags); |
| 1513 | void *IP = nullptr; |
| 1514 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1515 | return SDValue(E, 0); |
| 1516 | |
| 1517 | auto *N = newSDNode<TargetIndexSDNode>(Index, VT, Offset, TargetFlags); |
| 1518 | CSEMap.InsertNode(N, IP); |
| 1519 | InsertNode(N); |
| 1520 | return SDValue(N, 0); |
| 1521 | } |
| 1522 | |
| 1523 | SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) { |
| 1524 | FoldingSetNodeID ID; |
| 1525 | AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), None); |
| 1526 | ID.AddPointer(MBB); |
| 1527 | void *IP = nullptr; |
| 1528 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1529 | return SDValue(E, 0); |
| 1530 | |
| 1531 | auto *N = newSDNode<BasicBlockSDNode>(MBB); |
| 1532 | CSEMap.InsertNode(N, IP); |
| 1533 | InsertNode(N); |
| 1534 | return SDValue(N, 0); |
| 1535 | } |
| 1536 | |
| 1537 | SDValue SelectionDAG::getValueType(EVT VT) { |
| 1538 | if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >= |
| 1539 | ValueTypeNodes.size()) |
| 1540 | ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1); |
| 1541 | |
| 1542 | SDNode *&N = VT.isExtended() ? |
| 1543 | ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy]; |
| 1544 | |
| 1545 | if (N) return SDValue(N, 0); |
| 1546 | N = newSDNode<VTSDNode>(VT); |
| 1547 | InsertNode(N); |
| 1548 | return SDValue(N, 0); |
| 1549 | } |
| 1550 | |
| 1551 | SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) { |
| 1552 | SDNode *&N = ExternalSymbols[Sym]; |
| 1553 | if (N) return SDValue(N, 0); |
| 1554 | N = newSDNode<ExternalSymbolSDNode>(false, Sym, 0, VT); |
| 1555 | InsertNode(N); |
| 1556 | return SDValue(N, 0); |
| 1557 | } |
| 1558 | |
| 1559 | SDValue SelectionDAG::getExternalFunctionSymbol(const char *Sym) { |
| 1560 | auto AddrSpace = getDataLayout().getProgramAddressSpace(); |
| 1561 | return getExternalSymbol(Sym, TLI->getPointerTy(getDataLayout(), AddrSpace)); |
| 1562 | } |
| 1563 | |
| 1564 | SDValue SelectionDAG::getMCSymbol(MCSymbol *Sym, EVT VT) { |
| 1565 | SDNode *&N = MCSymbols[Sym]; |
| 1566 | if (N) |
| 1567 | return SDValue(N, 0); |
| 1568 | N = newSDNode<MCSymbolSDNode>(Sym, VT); |
| 1569 | InsertNode(N); |
| 1570 | return SDValue(N, 0); |
| 1571 | } |
| 1572 | |
| 1573 | SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT, |
| 1574 | unsigned char TargetFlags) { |
| 1575 | SDNode *&N = |
| 1576 | TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym, |
| 1577 | TargetFlags)]; |
| 1578 | if (N) return SDValue(N, 0); |
| 1579 | N = newSDNode<ExternalSymbolSDNode>(true, Sym, TargetFlags, VT); |
| 1580 | InsertNode(N); |
| 1581 | return SDValue(N, 0); |
| 1582 | } |
| 1583 | |
| 1584 | SDValue |
| 1585 | SelectionDAG::getTargetExternalFunctionSymbol(const char *Sym, |
| 1586 | unsigned char TargetFlags) { |
| 1587 | auto AddrSpace = getDataLayout().getProgramAddressSpace(); |
| 1588 | return getTargetExternalSymbol( |
| 1589 | Sym, TLI->getPointerTy(getDataLayout(), AddrSpace), TargetFlags); |
| 1590 | } |
| 1591 | |
| 1592 | SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) { |
| 1593 | if ((unsigned)Cond >= CondCodeNodes.size()) |
| 1594 | CondCodeNodes.resize(Cond+1); |
| 1595 | |
| 1596 | if (!CondCodeNodes[Cond]) { |
| 1597 | auto *N = newSDNode<CondCodeSDNode>(Cond); |
| 1598 | CondCodeNodes[Cond] = N; |
| 1599 | InsertNode(N); |
| 1600 | } |
| 1601 | |
| 1602 | return SDValue(CondCodeNodes[Cond], 0); |
| 1603 | } |
| 1604 | |
| 1605 | /// Swaps the values of N1 and N2. Swaps all indices in the shuffle mask M that |
| 1606 | /// point at N1 to point at N2 and indices that point at N2 to point at N1. |
| 1607 | static void commuteShuffle(SDValue &N1, SDValue &N2, MutableArrayRef<int> M) { |
| 1608 | std::swap(N1, N2); |
| 1609 | ShuffleVectorSDNode::commuteMask(M); |
| 1610 | } |
| 1611 | |
| 1612 | SDValue SelectionDAG::getVectorShuffle(EVT VT, const SDLoc &dl, SDValue N1, |
| 1613 | SDValue N2, ArrayRef<int> Mask) { |
| 1614 | assert(VT.getVectorNumElements() == Mask.size() && |
| 1615 | "Must have the same number of vector elements as mask elements!" ); |
| 1616 | assert(VT == N1.getValueType() && VT == N2.getValueType() && |
| 1617 | "Invalid VECTOR_SHUFFLE" ); |
| 1618 | |
| 1619 | // Canonicalize shuffle undef, undef -> undef |
| 1620 | if (N1.isUndef() && N2.isUndef()) |
| 1621 | return getUNDEF(VT); |
| 1622 | |
| 1623 | // Validate that all indices in Mask are within the range of the elements |
| 1624 | // input to the shuffle. |
| 1625 | int NElts = Mask.size(); |
| 1626 | assert(llvm::all_of(Mask, |
| 1627 | [&](int M) { return M < (NElts * 2) && M >= -1; }) && |
| 1628 | "Index out of range" ); |
| 1629 | |
| 1630 | // Copy the mask so we can do any needed cleanup. |
| 1631 | SmallVector<int, 8> MaskVec(Mask.begin(), Mask.end()); |
| 1632 | |
| 1633 | // Canonicalize shuffle v, v -> v, undef |
| 1634 | if (N1 == N2) { |
| 1635 | N2 = getUNDEF(VT); |
| 1636 | for (int i = 0; i != NElts; ++i) |
| 1637 | if (MaskVec[i] >= NElts) MaskVec[i] -= NElts; |
| 1638 | } |
| 1639 | |
| 1640 | // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask. |
| 1641 | if (N1.isUndef()) |
| 1642 | commuteShuffle(N1, N2, MaskVec); |
| 1643 | |
| 1644 | if (TLI->hasVectorBlend()) { |
| 1645 | // If shuffling a splat, try to blend the splat instead. We do this here so |
| 1646 | // that even when this arises during lowering we don't have to re-handle it. |
| 1647 | auto BlendSplat = [&](BuildVectorSDNode *BV, int Offset) { |
| 1648 | BitVector UndefElements; |
| 1649 | SDValue Splat = BV->getSplatValue(&UndefElements); |
| 1650 | if (!Splat) |
| 1651 | return; |
| 1652 | |
| 1653 | for (int i = 0; i < NElts; ++i) { |
| 1654 | if (MaskVec[i] < Offset || MaskVec[i] >= (Offset + NElts)) |
| 1655 | continue; |
| 1656 | |
| 1657 | // If this input comes from undef, mark it as such. |
| 1658 | if (UndefElements[MaskVec[i] - Offset]) { |
| 1659 | MaskVec[i] = -1; |
| 1660 | continue; |
| 1661 | } |
| 1662 | |
| 1663 | // If we can blend a non-undef lane, use that instead. |
| 1664 | if (!UndefElements[i]) |
| 1665 | MaskVec[i] = i + Offset; |
| 1666 | } |
| 1667 | }; |
| 1668 | if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1)) |
| 1669 | BlendSplat(N1BV, 0); |
| 1670 | if (auto *N2BV = dyn_cast<BuildVectorSDNode>(N2)) |
| 1671 | BlendSplat(N2BV, NElts); |
| 1672 | } |
| 1673 | |
| 1674 | // Canonicalize all index into lhs, -> shuffle lhs, undef |
| 1675 | // Canonicalize all index into rhs, -> shuffle rhs, undef |
| 1676 | bool AllLHS = true, AllRHS = true; |
| 1677 | bool N2Undef = N2.isUndef(); |
| 1678 | for (int i = 0; i != NElts; ++i) { |
| 1679 | if (MaskVec[i] >= NElts) { |
| 1680 | if (N2Undef) |
| 1681 | MaskVec[i] = -1; |
| 1682 | else |
| 1683 | AllLHS = false; |
| 1684 | } else if (MaskVec[i] >= 0) { |
| 1685 | AllRHS = false; |
| 1686 | } |
| 1687 | } |
| 1688 | if (AllLHS && AllRHS) |
| 1689 | return getUNDEF(VT); |
| 1690 | if (AllLHS && !N2Undef) |
| 1691 | N2 = getUNDEF(VT); |
| 1692 | if (AllRHS) { |
| 1693 | N1 = getUNDEF(VT); |
| 1694 | commuteShuffle(N1, N2, MaskVec); |
| 1695 | } |
| 1696 | // Reset our undef status after accounting for the mask. |
| 1697 | N2Undef = N2.isUndef(); |
| 1698 | // Re-check whether both sides ended up undef. |
| 1699 | if (N1.isUndef() && N2Undef) |
| 1700 | return getUNDEF(VT); |
| 1701 | |
| 1702 | // If Identity shuffle return that node. |
| 1703 | bool Identity = true, AllSame = true; |
| 1704 | for (int i = 0; i != NElts; ++i) { |
| 1705 | if (MaskVec[i] >= 0 && MaskVec[i] != i) Identity = false; |
| 1706 | if (MaskVec[i] != MaskVec[0]) AllSame = false; |
| 1707 | } |
| 1708 | if (Identity && NElts) |
| 1709 | return N1; |
| 1710 | |
| 1711 | // Shuffling a constant splat doesn't change the result. |
| 1712 | if (N2Undef) { |
| 1713 | SDValue V = N1; |
| 1714 | |
| 1715 | // Look through any bitcasts. We check that these don't change the number |
| 1716 | // (and size) of elements and just changes their types. |
| 1717 | while (V.getOpcode() == ISD::BITCAST) |
| 1718 | V = V->getOperand(0); |
| 1719 | |
| 1720 | // A splat should always show up as a build vector node. |
| 1721 | if (auto *BV = dyn_cast<BuildVectorSDNode>(V)) { |
| 1722 | BitVector UndefElements; |
| 1723 | SDValue Splat = BV->getSplatValue(&UndefElements); |
| 1724 | // If this is a splat of an undef, shuffling it is also undef. |
| 1725 | if (Splat && Splat.isUndef()) |
| 1726 | return getUNDEF(VT); |
| 1727 | |
| 1728 | bool = |
| 1729 | V.getValueType().getVectorNumElements() == VT.getVectorNumElements(); |
| 1730 | |
| 1731 | // We only have a splat which can skip shuffles if there is a splatted |
| 1732 | // value and no undef lanes rearranged by the shuffle. |
| 1733 | if (Splat && UndefElements.none()) { |
| 1734 | // Splat of <x, x, ..., x>, return <x, x, ..., x>, provided that the |
| 1735 | // number of elements match or the value splatted is a zero constant. |
| 1736 | if (SameNumElts) |
| 1737 | return N1; |
| 1738 | if (auto *C = dyn_cast<ConstantSDNode>(Splat)) |
| 1739 | if (C->isNullValue()) |
| 1740 | return N1; |
| 1741 | } |
| 1742 | |
| 1743 | // If the shuffle itself creates a splat, build the vector directly. |
| 1744 | if (AllSame && SameNumElts) { |
| 1745 | EVT BuildVT = BV->getValueType(0); |
| 1746 | const SDValue &Splatted = BV->getOperand(MaskVec[0]); |
| 1747 | SDValue NewBV = getSplatBuildVector(BuildVT, dl, Splatted); |
| 1748 | |
| 1749 | // We may have jumped through bitcasts, so the type of the |
| 1750 | // BUILD_VECTOR may not match the type of the shuffle. |
| 1751 | if (BuildVT != VT) |
| 1752 | NewBV = getNode(ISD::BITCAST, dl, VT, NewBV); |
| 1753 | return NewBV; |
| 1754 | } |
| 1755 | } |
| 1756 | } |
| 1757 | |
| 1758 | FoldingSetNodeID ID; |
| 1759 | SDValue Ops[2] = { N1, N2 }; |
| 1760 | AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops); |
| 1761 | for (int i = 0; i != NElts; ++i) |
| 1762 | ID.AddInteger(MaskVec[i]); |
| 1763 | |
| 1764 | void* IP = nullptr; |
| 1765 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) |
| 1766 | return SDValue(E, 0); |
| 1767 | |
| 1768 | // Allocate the mask array for the node out of the BumpPtrAllocator, since |
| 1769 | // SDNode doesn't have access to it. This memory will be "leaked" when |
| 1770 | // the node is deallocated, but recovered when the NodeAllocator is released. |
| 1771 | int *MaskAlloc = OperandAllocator.Allocate<int>(NElts); |
| 1772 | llvm::copy(MaskVec, MaskAlloc); |
| 1773 | |
| 1774 | auto *N = newSDNode<ShuffleVectorSDNode>(VT, dl.getIROrder(), |
| 1775 | dl.getDebugLoc(), MaskAlloc); |
| 1776 | createOperands(N, Ops); |
| 1777 | |
| 1778 | CSEMap.InsertNode(N, IP); |
| 1779 | InsertNode(N); |
| 1780 | SDValue V = SDValue(N, 0); |
| 1781 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 1782 | return V; |
| 1783 | } |
| 1784 | |
| 1785 | SDValue SelectionDAG::getCommutedVectorShuffle(const ShuffleVectorSDNode &SV) { |
| 1786 | EVT VT = SV.getValueType(0); |
| 1787 | SmallVector<int, 8> MaskVec(SV.getMask().begin(), SV.getMask().end()); |
| 1788 | ShuffleVectorSDNode::commuteMask(MaskVec); |
| 1789 | |
| 1790 | SDValue Op0 = SV.getOperand(0); |
| 1791 | SDValue Op1 = SV.getOperand(1); |
| 1792 | return getVectorShuffle(VT, SDLoc(&SV), Op1, Op0, MaskVec); |
| 1793 | } |
| 1794 | |
| 1795 | SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) { |
| 1796 | FoldingSetNodeID ID; |
| 1797 | AddNodeIDNode(ID, ISD::Register, getVTList(VT), None); |
| 1798 | ID.AddInteger(RegNo); |
| 1799 | void *IP = nullptr; |
| 1800 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1801 | return SDValue(E, 0); |
| 1802 | |
| 1803 | auto *N = newSDNode<RegisterSDNode>(RegNo, VT); |
| 1804 | N->SDNodeBits.IsDivergent = TLI->isSDNodeSourceOfDivergence(N, FLI, DA); |
| 1805 | CSEMap.InsertNode(N, IP); |
| 1806 | InsertNode(N); |
| 1807 | return SDValue(N, 0); |
| 1808 | } |
| 1809 | |
| 1810 | SDValue SelectionDAG::getRegisterMask(const uint32_t *RegMask) { |
| 1811 | FoldingSetNodeID ID; |
| 1812 | AddNodeIDNode(ID, ISD::RegisterMask, getVTList(MVT::Untyped), None); |
| 1813 | ID.AddPointer(RegMask); |
| 1814 | void *IP = nullptr; |
| 1815 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1816 | return SDValue(E, 0); |
| 1817 | |
| 1818 | auto *N = newSDNode<RegisterMaskSDNode>(RegMask); |
| 1819 | CSEMap.InsertNode(N, IP); |
| 1820 | InsertNode(N); |
| 1821 | return SDValue(N, 0); |
| 1822 | } |
| 1823 | |
| 1824 | SDValue SelectionDAG::getEHLabel(const SDLoc &dl, SDValue Root, |
| 1825 | MCSymbol *Label) { |
| 1826 | return getLabelNode(ISD::EH_LABEL, dl, Root, Label); |
| 1827 | } |
| 1828 | |
| 1829 | SDValue SelectionDAG::getLabelNode(unsigned Opcode, const SDLoc &dl, |
| 1830 | SDValue Root, MCSymbol *Label) { |
| 1831 | FoldingSetNodeID ID; |
| 1832 | SDValue Ops[] = { Root }; |
| 1833 | AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), Ops); |
| 1834 | ID.AddPointer(Label); |
| 1835 | void *IP = nullptr; |
| 1836 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1837 | return SDValue(E, 0); |
| 1838 | |
| 1839 | auto *N = |
| 1840 | newSDNode<LabelSDNode>(Opcode, dl.getIROrder(), dl.getDebugLoc(), Label); |
| 1841 | createOperands(N, Ops); |
| 1842 | |
| 1843 | CSEMap.InsertNode(N, IP); |
| 1844 | InsertNode(N); |
| 1845 | return SDValue(N, 0); |
| 1846 | } |
| 1847 | |
| 1848 | SDValue SelectionDAG::getBlockAddress(const BlockAddress *BA, EVT VT, |
| 1849 | int64_t Offset, |
| 1850 | bool isTarget, |
| 1851 | unsigned char TargetFlags) { |
| 1852 | unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress; |
| 1853 | |
| 1854 | FoldingSetNodeID ID; |
| 1855 | AddNodeIDNode(ID, Opc, getVTList(VT), None); |
| 1856 | ID.AddPointer(BA); |
| 1857 | ID.AddInteger(Offset); |
| 1858 | ID.AddInteger(TargetFlags); |
| 1859 | void *IP = nullptr; |
| 1860 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1861 | return SDValue(E, 0); |
| 1862 | |
| 1863 | auto *N = newSDNode<BlockAddressSDNode>(Opc, VT, BA, Offset, TargetFlags); |
| 1864 | CSEMap.InsertNode(N, IP); |
| 1865 | InsertNode(N); |
| 1866 | return SDValue(N, 0); |
| 1867 | } |
| 1868 | |
| 1869 | SDValue SelectionDAG::getSrcValue(const Value *V) { |
| 1870 | assert((!V || V->getType()->isPointerTy()) && |
| 1871 | "SrcValue is not a pointer?" ); |
| 1872 | |
| 1873 | FoldingSetNodeID ID; |
| 1874 | AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), None); |
| 1875 | ID.AddPointer(V); |
| 1876 | |
| 1877 | void *IP = nullptr; |
| 1878 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1879 | return SDValue(E, 0); |
| 1880 | |
| 1881 | auto *N = newSDNode<SrcValueSDNode>(V); |
| 1882 | CSEMap.InsertNode(N, IP); |
| 1883 | InsertNode(N); |
| 1884 | return SDValue(N, 0); |
| 1885 | } |
| 1886 | |
| 1887 | SDValue SelectionDAG::getMDNode(const MDNode *MD) { |
| 1888 | FoldingSetNodeID ID; |
| 1889 | AddNodeIDNode(ID, ISD::MDNODE_SDNODE, getVTList(MVT::Other), None); |
| 1890 | ID.AddPointer(MD); |
| 1891 | |
| 1892 | void *IP = nullptr; |
| 1893 | if (SDNode *E = FindNodeOrInsertPos(ID, IP)) |
| 1894 | return SDValue(E, 0); |
| 1895 | |
| 1896 | auto *N = newSDNode<MDNodeSDNode>(MD); |
| 1897 | CSEMap.InsertNode(N, IP); |
| 1898 | InsertNode(N); |
| 1899 | return SDValue(N, 0); |
| 1900 | } |
| 1901 | |
| 1902 | SDValue SelectionDAG::getBitcast(EVT VT, SDValue V) { |
| 1903 | if (VT == V.getValueType()) |
| 1904 | return V; |
| 1905 | |
| 1906 | return getNode(ISD::BITCAST, SDLoc(V), VT, V); |
| 1907 | } |
| 1908 | |
| 1909 | SDValue SelectionDAG::getAddrSpaceCast(const SDLoc &dl, EVT VT, SDValue Ptr, |
| 1910 | unsigned SrcAS, unsigned DestAS) { |
| 1911 | SDValue Ops[] = {Ptr}; |
| 1912 | FoldingSetNodeID ID; |
| 1913 | AddNodeIDNode(ID, ISD::ADDRSPACECAST, getVTList(VT), Ops); |
| 1914 | ID.AddInteger(SrcAS); |
| 1915 | ID.AddInteger(DestAS); |
| 1916 | |
| 1917 | void *IP = nullptr; |
| 1918 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) |
| 1919 | return SDValue(E, 0); |
| 1920 | |
| 1921 | auto *N = newSDNode<AddrSpaceCastSDNode>(dl.getIROrder(), dl.getDebugLoc(), |
| 1922 | VT, SrcAS, DestAS); |
| 1923 | createOperands(N, Ops); |
| 1924 | |
| 1925 | CSEMap.InsertNode(N, IP); |
| 1926 | InsertNode(N); |
| 1927 | return SDValue(N, 0); |
| 1928 | } |
| 1929 | |
| 1930 | /// getShiftAmountOperand - Return the specified value casted to |
| 1931 | /// the target's desired shift amount type. |
| 1932 | SDValue SelectionDAG::getShiftAmountOperand(EVT LHSTy, SDValue Op) { |
| 1933 | EVT OpTy = Op.getValueType(); |
| 1934 | EVT ShTy = TLI->getShiftAmountTy(LHSTy, getDataLayout()); |
| 1935 | if (OpTy == ShTy || OpTy.isVector()) return Op; |
| 1936 | |
| 1937 | return getZExtOrTrunc(Op, SDLoc(Op), ShTy); |
| 1938 | } |
| 1939 | |
| 1940 | SDValue SelectionDAG::expandVAArg(SDNode *Node) { |
| 1941 | SDLoc dl(Node); |
| 1942 | const TargetLowering &TLI = getTargetLoweringInfo(); |
| 1943 | const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); |
| 1944 | EVT VT = Node->getValueType(0); |
| 1945 | SDValue Tmp1 = Node->getOperand(0); |
| 1946 | SDValue Tmp2 = Node->getOperand(1); |
| 1947 | unsigned Align = Node->getConstantOperandVal(3); |
| 1948 | |
| 1949 | SDValue VAListLoad = getLoad( |
| 1950 | TLI.getPointerTy(getDataLayout(), getDataLayout().getAllocaAddrSpace()), |
| 1951 | dl, Tmp1, Tmp2, MachinePointerInfo(V)); |
| 1952 | SDValue VAList = VAListLoad; |
| 1953 | |
| 1954 | if (Align > TLI.getMinStackArgumentAlignment()) { |
| 1955 | assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2" ); |
| 1956 | |
| 1957 | VAList = getNode(ISD::ADD, dl, VAList.getValueType(), VAList, |
| 1958 | getConstant(Align - 1, dl, VAList.getValueType())); |
| 1959 | |
| 1960 | VAList = getNode(ISD::AND, dl, VAList.getValueType(), VAList, |
| 1961 | getConstant(-(int64_t)Align, dl, VAList.getValueType())); |
| 1962 | } |
| 1963 | |
| 1964 | // Increment the pointer, VAList, to the next vaarg |
| 1965 | Tmp1 = getNode(ISD::ADD, dl, VAList.getValueType(), VAList, |
| 1966 | getConstant(getDataLayout().getTypeAllocSize( |
| 1967 | VT.getTypeForEVT(*getContext())), |
| 1968 | dl, VAList.getValueType())); |
| 1969 | // Store the incremented VAList to the legalized pointer |
| 1970 | Tmp1 = |
| 1971 | getStore(VAListLoad.getValue(1), dl, Tmp1, Tmp2, MachinePointerInfo(V)); |
| 1972 | // Load the actual argument out of the pointer VAList |
| 1973 | return getLoad(VT, dl, Tmp1, VAList, MachinePointerInfo()); |
| 1974 | } |
| 1975 | |
| 1976 | SDValue SelectionDAG::expandVACopy(SDNode *Node) { |
| 1977 | SDLoc dl(Node); |
| 1978 | const TargetLowering &TLI = getTargetLoweringInfo(); |
| 1979 | // This defaults to loading a pointer from the input and storing it to the |
| 1980 | // output, returning the chain. |
| 1981 | const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue(); |
| 1982 | const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue(); |
| 1983 | SDValue Tmp1 = getLoad( |
| 1984 | TLI.getPointerTy(getDataLayout(), getDataLayout().getAllocaAddrSpace()), |
| 1985 | dl, Node->getOperand(0), Node->getOperand(2), MachinePointerInfo(VS)); |
| 1986 | return getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1), |
| 1987 | MachinePointerInfo(VD)); |
| 1988 | } |
| 1989 | |
| 1990 | SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) { |
| 1991 | MachineFrameInfo &MFI = getMachineFunction().getFrameInfo(); |
| 1992 | unsigned ByteSize = VT.getStoreSize(); |
| 1993 | Type *Ty = VT.getTypeForEVT(*getContext()); |
| 1994 | unsigned StackAlign = |
| 1995 | std::max((unsigned)getDataLayout().getPrefTypeAlignment(Ty), minAlign); |
| 1996 | |
| 1997 | int FrameIdx = MFI.CreateStackObject(ByteSize, StackAlign, false); |
| 1998 | return getFrameIndex(FrameIdx, TLI->getFrameIndexTy(getDataLayout())); |
| 1999 | } |
| 2000 | |
| 2001 | SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) { |
| 2002 | unsigned Bytes = std::max(VT1.getStoreSize(), VT2.getStoreSize()); |
| 2003 | Type *Ty1 = VT1.getTypeForEVT(*getContext()); |
| 2004 | Type *Ty2 = VT2.getTypeForEVT(*getContext()); |
| 2005 | const DataLayout &DL = getDataLayout(); |
| 2006 | unsigned Align = |
| 2007 | std::max(DL.getPrefTypeAlignment(Ty1), DL.getPrefTypeAlignment(Ty2)); |
| 2008 | |
| 2009 | MachineFrameInfo &MFI = getMachineFunction().getFrameInfo(); |
| 2010 | int FrameIdx = MFI.CreateStackObject(Bytes, Align, false); |
| 2011 | return getFrameIndex(FrameIdx, TLI->getFrameIndexTy(getDataLayout())); |
| 2012 | } |
| 2013 | |
| 2014 | SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1, SDValue N2, |
| 2015 | ISD::CondCode Cond, const SDLoc &dl) { |
| 2016 | EVT OpVT = N1.getValueType(); |
| 2017 | |
| 2018 | // These setcc operations always fold. |
| 2019 | switch (Cond) { |
| 2020 | default: break; |
| 2021 | case ISD::SETFALSE: |
| 2022 | case ISD::SETFALSE2: return getBoolConstant(false, dl, VT, OpVT); |
| 2023 | case ISD::SETTRUE: |
| 2024 | case ISD::SETTRUE2: return getBoolConstant(true, dl, VT, OpVT); |
| 2025 | |
| 2026 | case ISD::SETOEQ: |
| 2027 | case ISD::SETOGT: |
| 2028 | case ISD::SETOGE: |
| 2029 | case ISD::SETOLT: |
| 2030 | case ISD::SETOLE: |
| 2031 | case ISD::SETONE: |
| 2032 | case ISD::SETO: |
| 2033 | case ISD::SETUO: |
| 2034 | case ISD::SETUEQ: |
| 2035 | case ISD::SETUNE: |
| 2036 | assert(!OpVT.isInteger() && "Illegal setcc for integer!" ); |
| 2037 | break; |
| 2038 | } |
| 2039 | |
| 2040 | if (OpVT.isInteger()) { |
| 2041 | // For EQ and NE, we can always pick a value for the undef to make the |
| 2042 | // predicate pass or fail, so we can return undef. |
| 2043 | // Matches behavior in llvm::ConstantFoldCompareInstruction. |
| 2044 | // icmp eq/ne X, undef -> undef. |
| 2045 | if ((N1.isUndef() || N2.isUndef()) && |
| 2046 | (Cond == ISD::SETEQ || Cond == ISD::SETNE)) |
| 2047 | return getUNDEF(VT); |
| 2048 | |
| 2049 | // If both operands are undef, we can return undef for int comparison. |
| 2050 | // icmp undef, undef -> undef. |
| 2051 | if (N1.isUndef() && N2.isUndef()) |
| 2052 | return getUNDEF(VT); |
| 2053 | |
| 2054 | // icmp X, X -> true/false |
| 2055 | // icmp X, undef -> true/false because undef could be X. |
| 2056 | if (N1 == N2) |
| 2057 | return getBoolConstant(ISD::isTrueWhenEqual(Cond), dl, VT, OpVT); |
| 2058 | } |
| 2059 | |
| 2060 | if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2)) { |
| 2061 | const APInt &C2 = N2C->getAPIntValue(); |
| 2062 | if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1)) { |
| 2063 | const APInt &C1 = N1C->getAPIntValue(); |
| 2064 | |
| 2065 | switch (Cond) { |
| 2066 | default: llvm_unreachable("Unknown integer setcc!" ); |
| 2067 | case ISD::SETEQ: return getBoolConstant(C1 == C2, dl, VT, OpVT); |
| 2068 | case ISD::SETNE: return getBoolConstant(C1 != C2, dl, VT, OpVT); |
| 2069 | case ISD::SETULT: return getBoolConstant(C1.ult(C2), dl, VT, OpVT); |
| 2070 | case ISD::SETUGT: return getBoolConstant(C1.ugt(C2), dl, VT, OpVT); |
| 2071 | case ISD::SETULE: return getBoolConstant(C1.ule(C2), dl, VT, OpVT); |
| 2072 | case ISD::SETUGE: return getBoolConstant(C1.uge(C2), dl, VT, OpVT); |
| 2073 | case ISD::SETLT: return getBoolConstant(C1.slt(C2), dl, VT, OpVT); |
| 2074 | case ISD::SETGT: return getBoolConstant(C1.sgt(C2), dl, VT, OpVT); |
| 2075 | case ISD::SETLE: return getBoolConstant(C1.sle(C2), dl, VT, OpVT); |
| 2076 | case ISD::SETGE: return getBoolConstant(C1.sge(C2), dl, VT, OpVT); |
| 2077 | } |
| 2078 | } |
| 2079 | } |
| 2080 | |
| 2081 | auto *N1CFP = dyn_cast<ConstantFPSDNode>(N1); |
| 2082 | auto *N2CFP = dyn_cast<ConstantFPSDNode>(N2); |
| 2083 | |
| 2084 | if (N1CFP && N2CFP) { |
| 2085 | APFloat::cmpResult R = N1CFP->getValueAPF().compare(N2CFP->getValueAPF()); |
| 2086 | switch (Cond) { |
| 2087 | default: break; |
| 2088 | case ISD::SETEQ: if (R==APFloat::cmpUnordered) |
| 2089 | return getUNDEF(VT); |
| 2090 | LLVM_FALLTHROUGH; |
| 2091 | case ISD::SETOEQ: return getBoolConstant(R==APFloat::cmpEqual, dl, VT, |
| 2092 | OpVT); |
| 2093 | case ISD::SETNE: if (R==APFloat::cmpUnordered) |
| 2094 | return getUNDEF(VT); |
| 2095 | LLVM_FALLTHROUGH; |
| 2096 | case ISD::SETONE: return getBoolConstant(R==APFloat::cmpGreaterThan || |
| 2097 | R==APFloat::cmpLessThan, dl, VT, |
| 2098 | OpVT); |
| 2099 | case ISD::SETLT: if (R==APFloat::cmpUnordered) |
| 2100 | return getUNDEF(VT); |
| 2101 | LLVM_FALLTHROUGH; |
| 2102 | case ISD::SETOLT: return getBoolConstant(R==APFloat::cmpLessThan, dl, VT, |
| 2103 | OpVT); |
| 2104 | case ISD::SETGT: if (R==APFloat::cmpUnordered) |
| 2105 | return getUNDEF(VT); |
| 2106 | LLVM_FALLTHROUGH; |
| 2107 | case ISD::SETOGT: return getBoolConstant(R==APFloat::cmpGreaterThan, dl, |
| 2108 | VT, OpVT); |
| 2109 | case ISD::SETLE: if (R==APFloat::cmpUnordered) |
| 2110 | return getUNDEF(VT); |
| 2111 | LLVM_FALLTHROUGH; |
| 2112 | case ISD::SETOLE: return getBoolConstant(R==APFloat::cmpLessThan || |
| 2113 | R==APFloat::cmpEqual, dl, VT, |
| 2114 | OpVT); |
| 2115 | case ISD::SETGE: if (R==APFloat::cmpUnordered) |
| 2116 | return getUNDEF(VT); |
| 2117 | LLVM_FALLTHROUGH; |
| 2118 | case ISD::SETOGE: return getBoolConstant(R==APFloat::cmpGreaterThan || |
| 2119 | R==APFloat::cmpEqual, dl, VT, OpVT); |
| 2120 | case ISD::SETO: return getBoolConstant(R!=APFloat::cmpUnordered, dl, VT, |
| 2121 | OpVT); |
| 2122 | case ISD::SETUO: return getBoolConstant(R==APFloat::cmpUnordered, dl, VT, |
| 2123 | OpVT); |
| 2124 | case ISD::SETUEQ: return getBoolConstant(R==APFloat::cmpUnordered || |
| 2125 | R==APFloat::cmpEqual, dl, VT, |
| 2126 | OpVT); |
| 2127 | case ISD::SETUNE: return getBoolConstant(R!=APFloat::cmpEqual, dl, VT, |
| 2128 | OpVT); |
| 2129 | case ISD::SETULT: return getBoolConstant(R==APFloat::cmpUnordered || |
| 2130 | R==APFloat::cmpLessThan, dl, VT, |
| 2131 | OpVT); |
| 2132 | case ISD::SETUGT: return getBoolConstant(R==APFloat::cmpGreaterThan || |
| 2133 | R==APFloat::cmpUnordered, dl, VT, |
| 2134 | OpVT); |
| 2135 | case ISD::SETULE: return getBoolConstant(R!=APFloat::cmpGreaterThan, dl, |
| 2136 | VT, OpVT); |
| 2137 | case ISD::SETUGE: return getBoolConstant(R!=APFloat::cmpLessThan, dl, VT, |
| 2138 | OpVT); |
| 2139 | } |
| 2140 | } else if (N1CFP && OpVT.isSimple() && !N2.isUndef()) { |
| 2141 | // Ensure that the constant occurs on the RHS. |
| 2142 | ISD::CondCode SwappedCond = ISD::getSetCCSwappedOperands(Cond); |
| 2143 | if (!TLI->isCondCodeLegal(SwappedCond, OpVT.getSimpleVT())) |
| 2144 | return SDValue(); |
| 2145 | return getSetCC(dl, VT, N2, N1, SwappedCond); |
| 2146 | } else if ((N2CFP && N2CFP->getValueAPF().isNaN()) || |
| 2147 | (OpVT.isFloatingPoint() && (N1.isUndef() || N2.isUndef()))) { |
| 2148 | // If an operand is known to be a nan (or undef that could be a nan), we can |
| 2149 | // fold it. |
| 2150 | // Choosing NaN for the undef will always make unordered comparison succeed |
| 2151 | // and ordered comparison fails. |
| 2152 | // Matches behavior in llvm::ConstantFoldCompareInstruction. |
| 2153 | switch (ISD::getUnorderedFlavor(Cond)) { |
| 2154 | default: |
| 2155 | llvm_unreachable("Unknown flavor!" ); |
| 2156 | case 0: // Known false. |
| 2157 | return getBoolConstant(false, dl, VT, OpVT); |
| 2158 | case 1: // Known true. |
| 2159 | return getBoolConstant(true, dl, VT, OpVT); |
| 2160 | case 2: // Undefined. |
| 2161 | return getUNDEF(VT); |
| 2162 | } |
| 2163 | } |
| 2164 | |
| 2165 | // Could not fold it. |
| 2166 | return SDValue(); |
| 2167 | } |
| 2168 | |
| 2169 | /// See if the specified operand can be simplified with the knowledge that only |
| 2170 | /// the bits specified by DemandedBits are used. |
| 2171 | /// TODO: really we should be making this into the DAG equivalent of |
| 2172 | /// SimplifyMultipleUseDemandedBits and not generate any new nodes. |
| 2173 | SDValue SelectionDAG::GetDemandedBits(SDValue V, const APInt &DemandedBits) { |
| 2174 | EVT VT = V.getValueType(); |
| 2175 | APInt DemandedElts = VT.isVector() |
| 2176 | ? APInt::getAllOnesValue(VT.getVectorNumElements()) |
| 2177 | : APInt(1, 1); |
| 2178 | return GetDemandedBits(V, DemandedBits, DemandedElts); |
| 2179 | } |
| 2180 | |
| 2181 | /// See if the specified operand can be simplified with the knowledge that only |
| 2182 | /// the bits specified by DemandedBits are used in the elements specified by |
| 2183 | /// DemandedElts. |
| 2184 | /// TODO: really we should be making this into the DAG equivalent of |
| 2185 | /// SimplifyMultipleUseDemandedBits and not generate any new nodes. |
| 2186 | SDValue SelectionDAG::GetDemandedBits(SDValue V, const APInt &DemandedBits, |
| 2187 | const APInt &DemandedElts) { |
| 2188 | switch (V.getOpcode()) { |
| 2189 | default: |
| 2190 | break; |
| 2191 | case ISD::Constant: { |
| 2192 | auto *CV = cast<ConstantSDNode>(V.getNode()); |
| 2193 | assert(CV && "Const value should be ConstSDNode." ); |
| 2194 | const APInt &CVal = CV->getAPIntValue(); |
| 2195 | APInt NewVal = CVal & DemandedBits; |
| 2196 | if (NewVal != CVal) |
| 2197 | return getConstant(NewVal, SDLoc(V), V.getValueType()); |
| 2198 | break; |
| 2199 | } |
| 2200 | case ISD::OR: |
| 2201 | case ISD::XOR: |
| 2202 | // If the LHS or RHS don't contribute bits to the or, drop them. |
| 2203 | if (MaskedValueIsZero(V.getOperand(0), DemandedBits)) |
| 2204 | return V.getOperand(1); |
| 2205 | if (MaskedValueIsZero(V.getOperand(1), DemandedBits)) |
| 2206 | return V.getOperand(0); |
| 2207 | break; |
| 2208 | case ISD::SRL: |
| 2209 | // Only look at single-use SRLs. |
| 2210 | if (!V.getNode()->hasOneUse()) |
| 2211 | break; |
| 2212 | if (auto *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) { |
| 2213 | // See if we can recursively simplify the LHS. |
| 2214 | unsigned Amt = RHSC->getZExtValue(); |
| 2215 | |
| 2216 | // Watch out for shift count overflow though. |
| 2217 | if (Amt >= DemandedBits.getBitWidth()) |
| 2218 | break; |
| 2219 | APInt SrcDemandedBits = DemandedBits << Amt; |
| 2220 | if (SDValue SimplifyLHS = |
| 2221 | GetDemandedBits(V.getOperand(0), SrcDemandedBits)) |
| 2222 | return getNode(ISD::SRL, SDLoc(V), V.getValueType(), SimplifyLHS, |
| 2223 | V.getOperand(1)); |
| 2224 | } |
| 2225 | break; |
| 2226 | case ISD::AND: { |
| 2227 | // X & -1 -> X (ignoring bits which aren't demanded). |
| 2228 | // Also handle the case where masked out bits in X are known to be zero. |
| 2229 | if (ConstantSDNode *RHSC = isConstOrConstSplat(V.getOperand(1))) { |
| 2230 | const APInt &AndVal = RHSC->getAPIntValue(); |
| 2231 | if (DemandedBits.isSubsetOf(AndVal) || |
| 2232 | DemandedBits.isSubsetOf(computeKnownBits(V.getOperand(0)).Zero | |
| 2233 | AndVal)) |
| 2234 | return V.getOperand(0); |
| 2235 | } |
| 2236 | break; |
| 2237 | } |
| 2238 | case ISD::ANY_EXTEND: { |
| 2239 | SDValue Src = V.getOperand(0); |
| 2240 | unsigned SrcBitWidth = Src.getScalarValueSizeInBits(); |
| 2241 | // Being conservative here - only peek through if we only demand bits in the |
| 2242 | // non-extended source (even though the extended bits are technically |
| 2243 | // undef). |
| 2244 | if (DemandedBits.getActiveBits() > SrcBitWidth) |
| 2245 | break; |
| 2246 | APInt SrcDemandedBits = DemandedBits.trunc(SrcBitWidth); |
| 2247 | if (SDValue DemandedSrc = GetDemandedBits(Src, SrcDemandedBits)) |
| 2248 | return getNode(ISD::ANY_EXTEND, SDLoc(V), V.getValueType(), DemandedSrc); |
| 2249 | break; |
| 2250 | } |
| 2251 | case ISD::SIGN_EXTEND_INREG: |
| 2252 | EVT ExVT = cast<VTSDNode>(V.getOperand(1))->getVT(); |
| 2253 | unsigned ExVTBits = ExVT.getScalarSizeInBits(); |
| 2254 | |
| 2255 | // If none of the extended bits are demanded, eliminate the sextinreg. |
| 2256 | if (DemandedBits.getActiveBits() <= ExVTBits) |
| 2257 | return V.getOperand(0); |
| 2258 | |
| 2259 | break; |
| 2260 | } |
| 2261 | return SDValue(); |
| 2262 | } |
| 2263 | |
| 2264 | /// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We |
| 2265 | /// use this predicate to simplify operations downstream. |
| 2266 | bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const { |
| 2267 | unsigned BitWidth = Op.getScalarValueSizeInBits(); |
| 2268 | return MaskedValueIsZero(Op, APInt::getSignMask(BitWidth), Depth); |
| 2269 | } |
| 2270 | |
| 2271 | /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use |
| 2272 | /// this predicate to simplify operations downstream. Mask is known to be zero |
| 2273 | /// for bits that V cannot have. |
| 2274 | bool SelectionDAG::MaskedValueIsZero(SDValue V, const APInt &Mask, |
| 2275 | unsigned Depth) const { |
| 2276 | EVT VT = V.getValueType(); |
| 2277 | APInt DemandedElts = VT.isVector() |
| 2278 | ? APInt::getAllOnesValue(VT.getVectorNumElements()) |
| 2279 | : APInt(1, 1); |
| 2280 | return MaskedValueIsZero(V, Mask, DemandedElts, Depth); |
| 2281 | } |
| 2282 | |
| 2283 | /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero in |
| 2284 | /// DemandedElts. We use this predicate to simplify operations downstream. |
| 2285 | /// Mask is known to be zero for bits that V cannot have. |
| 2286 | bool SelectionDAG::MaskedValueIsZero(SDValue V, const APInt &Mask, |
| 2287 | const APInt &DemandedElts, |
| 2288 | unsigned Depth) const { |
| 2289 | return Mask.isSubsetOf(computeKnownBits(V, DemandedElts, Depth).Zero); |
| 2290 | } |
| 2291 | |
| 2292 | /// isSplatValue - Return true if the vector V has the same value |
| 2293 | /// across all DemandedElts. |
| 2294 | bool SelectionDAG::isSplatValue(SDValue V, const APInt &DemandedElts, |
| 2295 | APInt &UndefElts) { |
| 2296 | if (!DemandedElts) |
| 2297 | return false; // No demanded elts, better to assume we don't know anything. |
| 2298 | |
| 2299 | EVT VT = V.getValueType(); |
| 2300 | assert(VT.isVector() && "Vector type expected" ); |
| 2301 | |
| 2302 | unsigned NumElts = VT.getVectorNumElements(); |
| 2303 | assert(NumElts == DemandedElts.getBitWidth() && "Vector size mismatch" ); |
| 2304 | UndefElts = APInt::getNullValue(NumElts); |
| 2305 | |
| 2306 | switch (V.getOpcode()) { |
| 2307 | case ISD::BUILD_VECTOR: { |
| 2308 | SDValue Scl; |
| 2309 | for (unsigned i = 0; i != NumElts; ++i) { |
| 2310 | SDValue Op = V.getOperand(i); |
| 2311 | if (Op.isUndef()) { |
| 2312 | UndefElts.setBit(i); |
| 2313 | continue; |
| 2314 | } |
| 2315 | if (!DemandedElts[i]) |
| 2316 | continue; |
| 2317 | if (Scl && Scl != Op) |
| 2318 | return false; |
| 2319 | Scl = Op; |
| 2320 | } |
| 2321 | return true; |
| 2322 | } |
| 2323 | case ISD::VECTOR_SHUFFLE: { |
| 2324 | // Check if this is a shuffle node doing a splat. |
| 2325 | // TODO: Do we need to handle shuffle(splat, undef, mask)? |
| 2326 | int SplatIndex = -1; |
| 2327 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(V)->getMask(); |
| 2328 | for (int i = 0; i != (int)NumElts; ++i) { |
| 2329 | int M = Mask[i]; |
| 2330 | if (M < 0) { |
| 2331 | UndefElts.setBit(i); |
| 2332 | continue; |
| 2333 | } |
| 2334 | if (!DemandedElts[i]) |
| 2335 | continue; |
| 2336 | if (0 <= SplatIndex && SplatIndex != M) |
| 2337 | return false; |
| 2338 | SplatIndex = M; |
| 2339 | } |
| 2340 | return true; |
| 2341 | } |
| 2342 | case ISD::EXTRACT_SUBVECTOR: { |
| 2343 | SDValue Src = V.getOperand(0); |
| 2344 | ConstantSDNode *SubIdx = dyn_cast<ConstantSDNode>(V.getOperand(1)); |
| 2345 | unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); |
| 2346 | if (SubIdx && SubIdx->getAPIntValue().ule(NumSrcElts - NumElts)) { |
| 2347 | // Offset the demanded elts by the subvector index. |
| 2348 | uint64_t Idx = SubIdx->getZExtValue(); |
| 2349 | APInt UndefSrcElts; |
| 2350 | APInt DemandedSrc = DemandedElts.zextOrSelf(NumSrcElts).shl(Idx); |
| 2351 | if (isSplatValue(Src, DemandedSrc, UndefSrcElts)) { |
| 2352 | UndefElts = UndefSrcElts.extractBits(NumElts, Idx); |
| 2353 | return true; |
| 2354 | } |
| 2355 | } |
| 2356 | break; |
| 2357 | } |
| 2358 | case ISD::ADD: |
| 2359 | case ISD::SUB: |
| 2360 | case ISD::AND: { |
| 2361 | APInt UndefLHS, UndefRHS; |
| 2362 | SDValue LHS = V.getOperand(0); |
| 2363 | SDValue RHS = V.getOperand(1); |
| 2364 | if (isSplatValue(LHS, DemandedElts, UndefLHS) && |
| 2365 | isSplatValue(RHS, DemandedElts, UndefRHS)) { |
| 2366 | UndefElts = UndefLHS | UndefRHS; |
| 2367 | return true; |
| 2368 | } |
| 2369 | break; |
| 2370 | } |
| 2371 | } |
| 2372 | |
| 2373 | return false; |
| 2374 | } |
| 2375 | |
| 2376 | /// Helper wrapper to main isSplatValue function. |
| 2377 | bool SelectionDAG::isSplatValue(SDValue V, bool AllowUndefs) { |
| 2378 | EVT VT = V.getValueType(); |
| 2379 | assert(VT.isVector() && "Vector type expected" ); |
| 2380 | unsigned NumElts = VT.getVectorNumElements(); |
| 2381 | |
| 2382 | APInt UndefElts; |
| 2383 | APInt DemandedElts = APInt::getAllOnesValue(NumElts); |
| 2384 | return isSplatValue(V, DemandedElts, UndefElts) && |
| 2385 | (AllowUndefs || !UndefElts); |
| 2386 | } |
| 2387 | |
| 2388 | SDValue SelectionDAG::getSplatSourceVector(SDValue V, int &SplatIdx) { |
| 2389 | V = peekThroughExtractSubvectors(V); |
| 2390 | |
| 2391 | EVT VT = V.getValueType(); |
| 2392 | unsigned Opcode = V.getOpcode(); |
| 2393 | switch (Opcode) { |
| 2394 | default: { |
| 2395 | APInt UndefElts; |
| 2396 | APInt DemandedElts = APInt::getAllOnesValue(VT.getVectorNumElements()); |
| 2397 | if (isSplatValue(V, DemandedElts, UndefElts)) { |
| 2398 | // Handle case where all demanded elements are UNDEF. |
| 2399 | if (DemandedElts.isSubsetOf(UndefElts)) { |
| 2400 | SplatIdx = 0; |
| 2401 | return getUNDEF(VT); |
| 2402 | } |
| 2403 | SplatIdx = (UndefElts & DemandedElts).countTrailingOnes(); |
| 2404 | return V; |
| 2405 | } |
| 2406 | break; |
| 2407 | } |
| 2408 | case ISD::VECTOR_SHUFFLE: { |
| 2409 | // Check if this is a shuffle node doing a splat. |
| 2410 | // TODO - remove this and rely purely on SelectionDAG::isSplatValue, |
| 2411 | // getTargetVShiftNode currently struggles without the splat source. |
| 2412 | auto *SVN = cast<ShuffleVectorSDNode>(V); |
| 2413 | if (!SVN->isSplat()) |
| 2414 | break; |
| 2415 | int Idx = SVN->getSplatIndex(); |
| 2416 | int NumElts = V.getValueType().getVectorNumElements(); |
| 2417 | SplatIdx = Idx % NumElts; |
| 2418 | return V.getOperand(Idx / NumElts); |
| 2419 | } |
| 2420 | } |
| 2421 | |
| 2422 | return SDValue(); |
| 2423 | } |
| 2424 | |
| 2425 | SDValue SelectionDAG::getSplatValue(SDValue V) { |
| 2426 | int SplatIdx; |
| 2427 | if (SDValue SrcVector = getSplatSourceVector(V, SplatIdx)) |
| 2428 | return getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(V), |
| 2429 | SrcVector.getValueType().getScalarType(), SrcVector, |
| 2430 | getIntPtrConstant(SplatIdx, SDLoc(V))); |
| 2431 | return SDValue(); |
| 2432 | } |
| 2433 | |
| 2434 | /// If a SHL/SRA/SRL node has a constant or splat constant shift amount that |
| 2435 | /// is less than the element bit-width of the shift node, return it. |
| 2436 | static const APInt *getValidShiftAmountConstant(SDValue V) { |
| 2437 | if (ConstantSDNode *SA = isConstOrConstSplat(V.getOperand(1))) { |
| 2438 | // Shifting more than the bitwidth is not valid. |
| 2439 | const APInt &ShAmt = SA->getAPIntValue(); |
| 2440 | if (ShAmt.ult(V.getScalarValueSizeInBits())) |
| 2441 | return &ShAmt; |
| 2442 | } |
| 2443 | return nullptr; |
| 2444 | } |
| 2445 | |
| 2446 | /// Determine which bits of Op are known to be either zero or one and return |
| 2447 | /// them in Known. For vectors, the known bits are those that are shared by |
| 2448 | /// every vector element. |
| 2449 | KnownBits SelectionDAG::computeKnownBits(SDValue Op, unsigned Depth) const { |
| 2450 | EVT VT = Op.getValueType(); |
| 2451 | APInt DemandedElts = VT.isVector() |
| 2452 | ? APInt::getAllOnesValue(VT.getVectorNumElements()) |
| 2453 | : APInt(1, 1); |
| 2454 | return computeKnownBits(Op, DemandedElts, Depth); |
| 2455 | } |
| 2456 | |
| 2457 | /// Determine which bits of Op are known to be either zero or one and return |
| 2458 | /// them in Known. The DemandedElts argument allows us to only collect the known |
| 2459 | /// bits that are shared by the requested vector elements. |
| 2460 | KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts, |
| 2461 | unsigned Depth) const { |
| 2462 | unsigned BitWidth = Op.getScalarValueSizeInBits(); |
| 2463 | |
| 2464 | KnownBits Known(BitWidth); // Don't know anything. |
| 2465 | |
| 2466 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { |
| 2467 | // We know all of the bits for a constant! |
| 2468 | Known.One = C->getAPIntValue(); |
| 2469 | Known.Zero = ~Known.One; |
| 2470 | return Known; |
| 2471 | } |
| 2472 | if (auto *C = dyn_cast<ConstantFPSDNode>(Op)) { |
| 2473 | // We know all of the bits for a constant fp! |
| 2474 | Known.One = C->getValueAPF().bitcastToAPInt(); |
| 2475 | Known.Zero = ~Known.One; |
| 2476 | return Known; |
| 2477 | } |
| 2478 | |
| 2479 | if (Depth == 6) |
| 2480 | return Known; // Limit search depth. |
| 2481 | |
| 2482 | KnownBits Known2; |
| 2483 | unsigned NumElts = DemandedElts.getBitWidth(); |
| 2484 | assert((!Op.getValueType().isVector() || |
| 2485 | NumElts == Op.getValueType().getVectorNumElements()) && |
| 2486 | "Unexpected vector size" ); |
| 2487 | |
| 2488 | if (!DemandedElts) |
| 2489 | return Known; // No demanded elts, better to assume we don't know anything. |
| 2490 | |
| 2491 | unsigned Opcode = Op.getOpcode(); |
| 2492 | switch (Opcode) { |
| 2493 | case ISD::BUILD_VECTOR: |
| 2494 | // Collect the known bits that are shared by every demanded vector element. |
| 2495 | Known.Zero.setAllBits(); Known.One.setAllBits(); |
| 2496 | for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) { |
| 2497 | if (!DemandedElts[i]) |
| 2498 | continue; |
| 2499 | |
| 2500 | SDValue SrcOp = Op.getOperand(i); |
| 2501 | Known2 = computeKnownBits(SrcOp, Depth + 1); |
| 2502 | |
| 2503 | // BUILD_VECTOR can implicitly truncate sources, we must handle this. |
| 2504 | if (SrcOp.getValueSizeInBits() != BitWidth) { |
| 2505 | assert(SrcOp.getValueSizeInBits() > BitWidth && |
| 2506 | "Expected BUILD_VECTOR implicit truncation" ); |
| 2507 | Known2 = Known2.trunc(BitWidth); |
| 2508 | } |
| 2509 | |
| 2510 | // Known bits are the values that are shared by every demanded element. |
| 2511 | Known.One &= Known2.One; |
| 2512 | Known.Zero &= Known2.Zero; |
| 2513 | |
| 2514 | // If we don't know any bits, early out. |
| 2515 | if (Known.isUnknown()) |
| 2516 | break; |
| 2517 | } |
| 2518 | break; |
| 2519 | case ISD::VECTOR_SHUFFLE: { |
| 2520 | // Collect the known bits that are shared by every vector element referenced |
| 2521 | // by the shuffle. |
| 2522 | APInt DemandedLHS(NumElts, 0), DemandedRHS(NumElts, 0); |
| 2523 | Known.Zero.setAllBits(); Known.One.setAllBits(); |
| 2524 | const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op); |
| 2525 | assert(NumElts == SVN->getMask().size() && "Unexpected vector size" ); |
| 2526 | for (unsigned i = 0; i != NumElts; ++i) { |
| 2527 | if (!DemandedElts[i]) |
| 2528 | continue; |
| 2529 | |
| 2530 | int M = SVN->getMaskElt(i); |
| 2531 | if (M < 0) { |
| 2532 | // For UNDEF elements, we don't know anything about the common state of |
| 2533 | // the shuffle result. |
| 2534 | Known.resetAll(); |
| 2535 | DemandedLHS.clearAllBits(); |
| 2536 | DemandedRHS.clearAllBits(); |
| 2537 | break; |
| 2538 | } |
| 2539 | |
| 2540 | if ((unsigned)M < NumElts) |
| 2541 | DemandedLHS.setBit((unsigned)M % NumElts); |
| 2542 | else |
| 2543 | DemandedRHS.setBit((unsigned)M % NumElts); |
| 2544 | } |
| 2545 | // Known bits are the values that are shared by every demanded element. |
| 2546 | if (!!DemandedLHS) { |
| 2547 | SDValue LHS = Op.getOperand(0); |
| 2548 | Known2 = computeKnownBits(LHS, DemandedLHS, Depth + 1); |
| 2549 | Known.One &= Known2.One; |
| 2550 | Known.Zero &= Known2.Zero; |
| 2551 | } |
| 2552 | // If we don't know any bits, early out. |
| 2553 | if (Known.isUnknown()) |
| 2554 | break; |
| 2555 | if (!!DemandedRHS) { |
| 2556 | SDValue RHS = Op.getOperand(1); |
| 2557 | Known2 = computeKnownBits(RHS, DemandedRHS, Depth + 1); |
| 2558 | Known.One &= Known2.One; |
| 2559 | Known.Zero &= Known2.Zero; |
| 2560 | } |
| 2561 | break; |
| 2562 | } |
| 2563 | case ISD::CONCAT_VECTORS: { |
| 2564 | // Split DemandedElts and test each of the demanded subvectors. |
| 2565 | Known.Zero.setAllBits(); Known.One.setAllBits(); |
| 2566 | EVT SubVectorVT = Op.getOperand(0).getValueType(); |
| 2567 | unsigned NumSubVectorElts = SubVectorVT.getVectorNumElements(); |
| 2568 | unsigned NumSubVectors = Op.getNumOperands(); |
| 2569 | for (unsigned i = 0; i != NumSubVectors; ++i) { |
| 2570 | APInt DemandedSub = DemandedElts.lshr(i * NumSubVectorElts); |
| 2571 | DemandedSub = DemandedSub.trunc(NumSubVectorElts); |
| 2572 | if (!!DemandedSub) { |
| 2573 | SDValue Sub = Op.getOperand(i); |
| 2574 | Known2 = computeKnownBits(Sub, DemandedSub, Depth + 1); |
| 2575 | Known.One &= Known2.One; |
| 2576 | Known.Zero &= Known2.Zero; |
| 2577 | } |
| 2578 | // If we don't know any bits, early out. |
| 2579 | if (Known.isUnknown()) |
| 2580 | break; |
| 2581 | } |
| 2582 | break; |
| 2583 | } |
| 2584 | case ISD::INSERT_SUBVECTOR: { |
| 2585 | // If we know the element index, demand any elements from the subvector and |
| 2586 | // the remainder from the src its inserted into, otherwise demand them all. |
| 2587 | SDValue Src = Op.getOperand(0); |
| 2588 | SDValue Sub = Op.getOperand(1); |
| 2589 | ConstantSDNode *SubIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2)); |
| 2590 | unsigned NumSubElts = Sub.getValueType().getVectorNumElements(); |
| 2591 | if (SubIdx && SubIdx->getAPIntValue().ule(NumElts - NumSubElts)) { |
| 2592 | Known.One.setAllBits(); |
| 2593 | Known.Zero.setAllBits(); |
| 2594 | uint64_t Idx = SubIdx->getZExtValue(); |
| 2595 | APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx); |
| 2596 | if (!!DemandedSubElts) { |
| 2597 | Known = computeKnownBits(Sub, DemandedSubElts, Depth + 1); |
| 2598 | if (Known.isUnknown()) |
| 2599 | break; // early-out. |
| 2600 | } |
| 2601 | APInt SubMask = APInt::getBitsSet(NumElts, Idx, Idx + NumSubElts); |
| 2602 | APInt DemandedSrcElts = DemandedElts & ~SubMask; |
| 2603 | if (!!DemandedSrcElts) { |
| 2604 | Known2 = computeKnownBits(Src, DemandedSrcElts, Depth + 1); |
| 2605 | Known.One &= Known2.One; |
| 2606 | Known.Zero &= Known2.Zero; |
| 2607 | } |
| 2608 | } else { |
| 2609 | Known = computeKnownBits(Sub, Depth + 1); |
| 2610 | if (Known.isUnknown()) |
| 2611 | break; // early-out. |
| 2612 | Known2 = computeKnownBits(Src, Depth + 1); |
| 2613 | Known.One &= Known2.One; |
| 2614 | Known.Zero &= Known2.Zero; |
| 2615 | } |
| 2616 | break; |
| 2617 | } |
| 2618 | case ISD::EXTRACT_SUBVECTOR: { |
| 2619 | // If we know the element index, just demand that subvector elements, |
| 2620 | // otherwise demand them all. |
| 2621 | SDValue Src = Op.getOperand(0); |
| 2622 | ConstantSDNode *SubIdx = dyn_cast<ConstantSDNode>(Op.getOperand(1)); |
| 2623 | unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); |
| 2624 | if (SubIdx && SubIdx->getAPIntValue().ule(NumSrcElts - NumElts)) { |
| 2625 | // Offset the demanded elts by the subvector index. |
| 2626 | uint64_t Idx = SubIdx->getZExtValue(); |
| 2627 | APInt DemandedSrc = DemandedElts.zextOrSelf(NumSrcElts).shl(Idx); |
| 2628 | Known = computeKnownBits(Src, DemandedSrc, Depth + 1); |
| 2629 | } else { |
| 2630 | Known = computeKnownBits(Src, Depth + 1); |
| 2631 | } |
| 2632 | break; |
| 2633 | } |
| 2634 | case ISD::SCALAR_TO_VECTOR: { |
| 2635 | // We know about scalar_to_vector as much as we know about it source, |
| 2636 | // which becomes the first element of otherwise unknown vector. |
| 2637 | if (DemandedElts != 1) |
| 2638 | break; |
| 2639 | |
| 2640 | SDValue N0 = Op.getOperand(0); |
| 2641 | Known = computeKnownBits(N0, Depth + 1); |
| 2642 | if (N0.getValueSizeInBits() != BitWidth) |
| 2643 | Known = Known.trunc(BitWidth); |
| 2644 | |
| 2645 | break; |
| 2646 | } |
| 2647 | case ISD::BITCAST: { |
| 2648 | SDValue N0 = Op.getOperand(0); |
| 2649 | EVT SubVT = N0.getValueType(); |
| 2650 | unsigned SubBitWidth = SubVT.getScalarSizeInBits(); |
| 2651 | |
| 2652 | // Ignore bitcasts from unsupported types. |
| 2653 | if (!(SubVT.isInteger() || SubVT.isFloatingPoint())) |
| 2654 | break; |
| 2655 | |
| 2656 | // Fast handling of 'identity' bitcasts. |
| 2657 | if (BitWidth == SubBitWidth) { |
| 2658 | Known = computeKnownBits(N0, DemandedElts, Depth + 1); |
| 2659 | break; |
| 2660 | } |
| 2661 | |
| 2662 | bool IsLE = getDataLayout().isLittleEndian(); |
| 2663 | |
| 2664 | // Bitcast 'small element' vector to 'large element' scalar/vector. |
| 2665 | if ((BitWidth % SubBitWidth) == 0) { |
| 2666 | assert(N0.getValueType().isVector() && "Expected bitcast from vector" ); |
| 2667 | |
| 2668 | // Collect known bits for the (larger) output by collecting the known |
| 2669 | // bits from each set of sub elements and shift these into place. |
| 2670 | // We need to separately call computeKnownBits for each set of |
| 2671 | // sub elements as the knownbits for each is likely to be different. |
| 2672 | unsigned SubScale = BitWidth / SubBitWidth; |
| 2673 | APInt SubDemandedElts(NumElts * SubScale, 0); |
| 2674 | for (unsigned i = 0; i != NumElts; ++i) |
| 2675 | if (DemandedElts[i]) |
| 2676 | SubDemandedElts.setBit(i * SubScale); |
| 2677 | |
| 2678 | for (unsigned i = 0; i != SubScale; ++i) { |
| 2679 | Known2 = computeKnownBits(N0, SubDemandedElts.shl(i), |
| 2680 | Depth + 1); |
| 2681 | unsigned Shifts = IsLE ? i : SubScale - 1 - i; |
| 2682 | Known.One |= Known2.One.zext(BitWidth).shl(SubBitWidth * Shifts); |
| 2683 | Known.Zero |= Known2.Zero.zext(BitWidth).shl(SubBitWidth * Shifts); |
| 2684 | } |
| 2685 | } |
| 2686 | |
| 2687 | // Bitcast 'large element' scalar/vector to 'small element' vector. |
| 2688 | if ((SubBitWidth % BitWidth) == 0) { |
| 2689 | assert(Op.getValueType().isVector() && "Expected bitcast to vector" ); |
| 2690 | |
| 2691 | // Collect known bits for the (smaller) output by collecting the known |
| 2692 | // bits from the overlapping larger input elements and extracting the |
| 2693 | // sub sections we actually care about. |
| 2694 | unsigned SubScale = SubBitWidth / BitWidth; |
| 2695 | APInt SubDemandedElts(NumElts / SubScale, 0); |
| 2696 | for (unsigned i = 0; i != NumElts; ++i) |
| 2697 | if (DemandedElts[i]) |
| 2698 | SubDemandedElts.setBit(i / SubScale); |
| 2699 | |
| 2700 | Known2 = computeKnownBits(N0, SubDemandedElts, Depth + 1); |
| 2701 | |
| 2702 | Known.Zero.setAllBits(); Known.One.setAllBits(); |
| 2703 | for (unsigned i = 0; i != NumElts; ++i) |
| 2704 | if (DemandedElts[i]) { |
| 2705 | unsigned Shifts = IsLE ? i : NumElts - 1 - i; |
| 2706 | unsigned Offset = (Shifts % SubScale) * BitWidth; |
| 2707 | Known.One &= Known2.One.lshr(Offset).trunc(BitWidth); |
| 2708 | Known.Zero &= Known2.Zero.lshr(Offset).trunc(BitWidth); |
| 2709 | // If we don't know any bits, early out. |
| 2710 | if (Known.isUnknown()) |
| 2711 | break; |
| 2712 | } |
| 2713 | } |
| 2714 | break; |
| 2715 | } |
| 2716 | case ISD::AND: |
| 2717 | // If either the LHS or the RHS are Zero, the result is zero. |
| 2718 | Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 2719 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2720 | |
| 2721 | // Output known-1 bits are only known if set in both the LHS & RHS. |
| 2722 | Known.One &= Known2.One; |
| 2723 | // Output known-0 are known to be clear if zero in either the LHS | RHS. |
| 2724 | Known.Zero |= Known2.Zero; |
| 2725 | break; |
| 2726 | case ISD::OR: |
| 2727 | Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 2728 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2729 | |
| 2730 | // Output known-0 bits are only known if clear in both the LHS & RHS. |
| 2731 | Known.Zero &= Known2.Zero; |
| 2732 | // Output known-1 are known to be set if set in either the LHS | RHS. |
| 2733 | Known.One |= Known2.One; |
| 2734 | break; |
| 2735 | case ISD::XOR: { |
| 2736 | Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 2737 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2738 | |
| 2739 | // Output known-0 bits are known if clear or set in both the LHS & RHS. |
| 2740 | APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One); |
| 2741 | // Output known-1 are known to be set if set in only one of the LHS, RHS. |
| 2742 | Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero); |
| 2743 | Known.Zero = KnownZeroOut; |
| 2744 | break; |
| 2745 | } |
| 2746 | case ISD::MUL: { |
| 2747 | Known = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 2748 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2749 | |
| 2750 | // If low bits are zero in either operand, output low known-0 bits. |
| 2751 | // Also compute a conservative estimate for high known-0 bits. |
| 2752 | // More trickiness is possible, but this is sufficient for the |
| 2753 | // interesting case of alignment computation. |
| 2754 | unsigned TrailZ = Known.countMinTrailingZeros() + |
| 2755 | Known2.countMinTrailingZeros(); |
| 2756 | unsigned LeadZ = std::max(Known.countMinLeadingZeros() + |
| 2757 | Known2.countMinLeadingZeros(), |
| 2758 | BitWidth) - BitWidth; |
| 2759 | |
| 2760 | Known.resetAll(); |
| 2761 | Known.Zero.setLowBits(std::min(TrailZ, BitWidth)); |
| 2762 | Known.Zero.setHighBits(std::min(LeadZ, BitWidth)); |
| 2763 | break; |
| 2764 | } |
| 2765 | case ISD::UDIV: { |
| 2766 | // For the purposes of computing leading zeros we can conservatively |
| 2767 | // treat a udiv as a logical right shift by the power of 2 known to |
| 2768 | // be less than the denominator. |
| 2769 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2770 | unsigned LeadZ = Known2.countMinLeadingZeros(); |
| 2771 | |
| 2772 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 2773 | unsigned RHSMaxLeadingZeros = Known2.countMaxLeadingZeros(); |
| 2774 | if (RHSMaxLeadingZeros != BitWidth) |
| 2775 | LeadZ = std::min(BitWidth, LeadZ + BitWidth - RHSMaxLeadingZeros - 1); |
| 2776 | |
| 2777 | Known.Zero.setHighBits(LeadZ); |
| 2778 | break; |
| 2779 | } |
| 2780 | case ISD::SELECT: |
| 2781 | case ISD::VSELECT: |
| 2782 | Known = computeKnownBits(Op.getOperand(2), DemandedElts, Depth+1); |
| 2783 | // If we don't know any bits, early out. |
| 2784 | if (Known.isUnknown()) |
| 2785 | break; |
| 2786 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth+1); |
| 2787 | |
| 2788 | // Only known if known in both the LHS and RHS. |
| 2789 | Known.One &= Known2.One; |
| 2790 | Known.Zero &= Known2.Zero; |
| 2791 | break; |
| 2792 | case ISD::SELECT_CC: |
| 2793 | Known = computeKnownBits(Op.getOperand(3), DemandedElts, Depth+1); |
| 2794 | // If we don't know any bits, early out. |
| 2795 | if (Known.isUnknown()) |
| 2796 | break; |
| 2797 | Known2 = computeKnownBits(Op.getOperand(2), DemandedElts, Depth+1); |
| 2798 | |
| 2799 | // Only known if known in both the LHS and RHS. |
| 2800 | Known.One &= Known2.One; |
| 2801 | Known.Zero &= Known2.Zero; |
| 2802 | break; |
| 2803 | case ISD::SMULO: |
| 2804 | case ISD::UMULO: |
| 2805 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: |
| 2806 | if (Op.getResNo() != 1) |
| 2807 | break; |
| 2808 | // The boolean result conforms to getBooleanContents. |
| 2809 | // If we know the result of a setcc has the top bits zero, use this info. |
| 2810 | // We know that we have an integer-based boolean since these operations |
| 2811 | // are only available for integer. |
| 2812 | if (TLI->getBooleanContents(Op.getValueType().isVector(), false) == |
| 2813 | TargetLowering::ZeroOrOneBooleanContent && |
| 2814 | BitWidth > 1) |
| 2815 | Known.Zero.setBitsFrom(1); |
| 2816 | break; |
| 2817 | case ISD::SETCC: |
| 2818 | // If we know the result of a setcc has the top bits zero, use this info. |
| 2819 | if (TLI->getBooleanContents(Op.getOperand(0).getValueType()) == |
| 2820 | TargetLowering::ZeroOrOneBooleanContent && |
| 2821 | BitWidth > 1) |
| 2822 | Known.Zero.setBitsFrom(1); |
| 2823 | break; |
| 2824 | case ISD::SHL: |
| 2825 | if (const APInt *ShAmt = getValidShiftAmountConstant(Op)) { |
| 2826 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2827 | unsigned Shift = ShAmt->getZExtValue(); |
| 2828 | Known.Zero <<= Shift; |
| 2829 | Known.One <<= Shift; |
| 2830 | // Low bits are known zero. |
| 2831 | Known.Zero.setLowBits(Shift); |
| 2832 | } |
| 2833 | break; |
| 2834 | case ISD::SRL: |
| 2835 | if (const APInt *ShAmt = getValidShiftAmountConstant(Op)) { |
| 2836 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2837 | unsigned Shift = ShAmt->getZExtValue(); |
| 2838 | Known.Zero.lshrInPlace(Shift); |
| 2839 | Known.One.lshrInPlace(Shift); |
| 2840 | // High bits are known zero. |
| 2841 | Known.Zero.setHighBits(Shift); |
| 2842 | } else if (auto *BV = dyn_cast<BuildVectorSDNode>(Op.getOperand(1))) { |
| 2843 | // If the shift amount is a vector of constants see if we can bound |
| 2844 | // the number of upper zero bits. |
| 2845 | unsigned ShiftAmountMin = BitWidth; |
| 2846 | for (unsigned i = 0; i != BV->getNumOperands(); ++i) { |
| 2847 | if (auto *C = dyn_cast<ConstantSDNode>(BV->getOperand(i))) { |
| 2848 | const APInt &ShAmt = C->getAPIntValue(); |
| 2849 | if (ShAmt.ult(BitWidth)) { |
| 2850 | ShiftAmountMin = std::min<unsigned>(ShiftAmountMin, |
| 2851 | ShAmt.getZExtValue()); |
| 2852 | continue; |
| 2853 | } |
| 2854 | } |
| 2855 | // Don't know anything. |
| 2856 | ShiftAmountMin = 0; |
| 2857 | break; |
| 2858 | } |
| 2859 | |
| 2860 | Known.Zero.setHighBits(ShiftAmountMin); |
| 2861 | } |
| 2862 | break; |
| 2863 | case ISD::SRA: |
| 2864 | if (const APInt *ShAmt = getValidShiftAmountConstant(Op)) { |
| 2865 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2866 | unsigned Shift = ShAmt->getZExtValue(); |
| 2867 | // Sign extend known zero/one bit (else is unknown). |
| 2868 | Known.Zero.ashrInPlace(Shift); |
| 2869 | Known.One.ashrInPlace(Shift); |
| 2870 | } |
| 2871 | break; |
| 2872 | case ISD::FSHL: |
| 2873 | case ISD::FSHR: |
| 2874 | if (ConstantSDNode *C = isConstOrConstSplat(Op.getOperand(2), DemandedElts)) { |
| 2875 | unsigned Amt = C->getAPIntValue().urem(BitWidth); |
| 2876 | |
| 2877 | // For fshl, 0-shift returns the 1st arg. |
| 2878 | // For fshr, 0-shift returns the 2nd arg. |
| 2879 | if (Amt == 0) { |
| 2880 | Known = computeKnownBits(Op.getOperand(Opcode == ISD::FSHL ? 0 : 1), |
| 2881 | DemandedElts, Depth + 1); |
| 2882 | break; |
| 2883 | } |
| 2884 | |
| 2885 | // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW))) |
| 2886 | // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW)) |
| 2887 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2888 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 2889 | if (Opcode == ISD::FSHL) { |
| 2890 | Known.One <<= Amt; |
| 2891 | Known.Zero <<= Amt; |
| 2892 | Known2.One.lshrInPlace(BitWidth - Amt); |
| 2893 | Known2.Zero.lshrInPlace(BitWidth - Amt); |
| 2894 | } else { |
| 2895 | Known.One <<= BitWidth - Amt; |
| 2896 | Known.Zero <<= BitWidth - Amt; |
| 2897 | Known2.One.lshrInPlace(Amt); |
| 2898 | Known2.Zero.lshrInPlace(Amt); |
| 2899 | } |
| 2900 | Known.One |= Known2.One; |
| 2901 | Known.Zero |= Known2.Zero; |
| 2902 | } |
| 2903 | break; |
| 2904 | case ISD::SIGN_EXTEND_INREG: { |
| 2905 | EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); |
| 2906 | unsigned EBits = EVT.getScalarSizeInBits(); |
| 2907 | |
| 2908 | // Sign extension. Compute the demanded bits in the result that are not |
| 2909 | // present in the input. |
| 2910 | APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits); |
| 2911 | |
| 2912 | APInt InSignMask = APInt::getSignMask(EBits); |
| 2913 | APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth, EBits); |
| 2914 | |
| 2915 | // If the sign extended bits are demanded, we know that the sign |
| 2916 | // bit is demanded. |
| 2917 | InSignMask = InSignMask.zext(BitWidth); |
| 2918 | if (NewBits.getBoolValue()) |
| 2919 | InputDemandedBits |= InSignMask; |
| 2920 | |
| 2921 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2922 | Known.One &= InputDemandedBits; |
| 2923 | Known.Zero &= InputDemandedBits; |
| 2924 | |
| 2925 | // If the sign bit of the input is known set or clear, then we know the |
| 2926 | // top bits of the result. |
| 2927 | if (Known.Zero.intersects(InSignMask)) { // Input sign bit known clear |
| 2928 | Known.Zero |= NewBits; |
| 2929 | Known.One &= ~NewBits; |
| 2930 | } else if (Known.One.intersects(InSignMask)) { // Input sign bit known set |
| 2931 | Known.One |= NewBits; |
| 2932 | Known.Zero &= ~NewBits; |
| 2933 | } else { // Input sign bit unknown |
| 2934 | Known.Zero &= ~NewBits; |
| 2935 | Known.One &= ~NewBits; |
| 2936 | } |
| 2937 | break; |
| 2938 | } |
| 2939 | case ISD::CTTZ: |
| 2940 | case ISD::CTTZ_ZERO_UNDEF: { |
| 2941 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2942 | // If we have a known 1, its position is our upper bound. |
| 2943 | unsigned PossibleTZ = Known2.countMaxTrailingZeros(); |
| 2944 | unsigned LowBits = Log2_32(PossibleTZ) + 1; |
| 2945 | Known.Zero.setBitsFrom(LowBits); |
| 2946 | break; |
| 2947 | } |
| 2948 | case ISD::CTLZ: |
| 2949 | case ISD::CTLZ_ZERO_UNDEF: { |
| 2950 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2951 | // If we have a known 1, its position is our upper bound. |
| 2952 | unsigned PossibleLZ = Known2.countMaxLeadingZeros(); |
| 2953 | unsigned LowBits = Log2_32(PossibleLZ) + 1; |
| 2954 | Known.Zero.setBitsFrom(LowBits); |
| 2955 | break; |
| 2956 | } |
| 2957 | case ISD::CTPOP: { |
| 2958 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 2959 | // If we know some of the bits are zero, they can't be one. |
| 2960 | unsigned PossibleOnes = Known2.countMaxPopulation(); |
| 2961 | Known.Zero.setBitsFrom(Log2_32(PossibleOnes) + 1); |
| 2962 | break; |
| 2963 | } |
| 2964 | case ISD::LOAD: { |
| 2965 | LoadSDNode *LD = cast<LoadSDNode>(Op); |
| 2966 | const Constant *Cst = TLI->getTargetConstantFromLoad(LD); |
| 2967 | if (ISD::isNON_EXTLoad(LD) && Cst) { |
| 2968 | // Determine any common known bits from the loaded constant pool value. |
| 2969 | Type *CstTy = Cst->getType(); |
| 2970 | if ((NumElts * BitWidth) == CstTy->getPrimitiveSizeInBits()) { |
| 2971 | // If its a vector splat, then we can (quickly) reuse the scalar path. |
| 2972 | // NOTE: We assume all elements match and none are UNDEF. |
| 2973 | if (CstTy->isVectorTy()) { |
| 2974 | if (const Constant *Splat = Cst->getSplatValue()) { |
| 2975 | Cst = Splat; |
| 2976 | CstTy = Cst->getType(); |
| 2977 | } |
| 2978 | } |
| 2979 | // TODO - do we need to handle different bitwidths? |
| 2980 | if (CstTy->isVectorTy() && BitWidth == CstTy->getScalarSizeInBits()) { |
| 2981 | // Iterate across all vector elements finding common known bits. |
| 2982 | Known.One.setAllBits(); |
| 2983 | Known.Zero.setAllBits(); |
| 2984 | for (unsigned i = 0; i != NumElts; ++i) { |
| 2985 | if (!DemandedElts[i]) |
| 2986 | continue; |
| 2987 | if (Constant *Elt = Cst->getAggregateElement(i)) { |
| 2988 | if (auto *CInt = dyn_cast<ConstantInt>(Elt)) { |
| 2989 | const APInt &Value = CInt->getValue(); |
| 2990 | Known.One &= Value; |
| 2991 | Known.Zero &= ~Value; |
| 2992 | continue; |
| 2993 | } |
| 2994 | if (auto *CFP = dyn_cast<ConstantFP>(Elt)) { |
| 2995 | APInt Value = CFP->getValueAPF().bitcastToAPInt(); |
| 2996 | Known.One &= Value; |
| 2997 | Known.Zero &= ~Value; |
| 2998 | continue; |
| 2999 | } |
| 3000 | } |
| 3001 | Known.One.clearAllBits(); |
| 3002 | Known.Zero.clearAllBits(); |
| 3003 | break; |
| 3004 | } |
| 3005 | } else if (BitWidth == CstTy->getPrimitiveSizeInBits()) { |
| 3006 | if (auto *CInt = dyn_cast<ConstantInt>(Cst)) { |
| 3007 | const APInt &Value = CInt->getValue(); |
| 3008 | Known.One = Value; |
| 3009 | Known.Zero = ~Value; |
| 3010 | } else if (auto *CFP = dyn_cast<ConstantFP>(Cst)) { |
| 3011 | APInt Value = CFP->getValueAPF().bitcastToAPInt(); |
| 3012 | Known.One = Value; |
| 3013 | Known.Zero = ~Value; |
| 3014 | } |
| 3015 | } |
| 3016 | } |
| 3017 | } else if (ISD::isZEXTLoad(Op.getNode()) && Op.getResNo() == 0) { |
| 3018 | // If this is a ZEXTLoad and we are looking at the loaded value. |
| 3019 | EVT VT = LD->getMemoryVT(); |
| 3020 | unsigned MemBits = VT.getScalarSizeInBits(); |
| 3021 | Known.Zero.setBitsFrom(MemBits); |
| 3022 | } else if (const MDNode *Ranges = LD->getRanges()) { |
| 3023 | if (LD->getExtensionType() == ISD::NON_EXTLOAD) |
| 3024 | computeKnownBitsFromRangeMetadata(*Ranges, Known); |
| 3025 | } |
| 3026 | break; |
| 3027 | } |
| 3028 | case ISD::ZERO_EXTEND_VECTOR_INREG: { |
| 3029 | EVT InVT = Op.getOperand(0).getValueType(); |
| 3030 | APInt InDemandedElts = DemandedElts.zextOrSelf(InVT.getVectorNumElements()); |
| 3031 | Known = computeKnownBits(Op.getOperand(0), InDemandedElts, Depth + 1); |
| 3032 | Known = Known.zext(BitWidth, true /* ExtendedBitsAreKnownZero */); |
| 3033 | break; |
| 3034 | } |
| 3035 | case ISD::ZERO_EXTEND: { |
| 3036 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3037 | Known = Known.zext(BitWidth, true /* ExtendedBitsAreKnownZero */); |
| 3038 | break; |
| 3039 | } |
| 3040 | case ISD::SIGN_EXTEND_VECTOR_INREG: { |
| 3041 | EVT InVT = Op.getOperand(0).getValueType(); |
| 3042 | APInt InDemandedElts = DemandedElts.zextOrSelf(InVT.getVectorNumElements()); |
| 3043 | Known = computeKnownBits(Op.getOperand(0), InDemandedElts, Depth + 1); |
| 3044 | // If the sign bit is known to be zero or one, then sext will extend |
| 3045 | // it to the top bits, else it will just zext. |
| 3046 | Known = Known.sext(BitWidth); |
| 3047 | break; |
| 3048 | } |
| 3049 | case ISD::SIGN_EXTEND: { |
| 3050 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3051 | // If the sign bit is known to be zero or one, then sext will extend |
| 3052 | // it to the top bits, else it will just zext. |
| 3053 | Known = Known.sext(BitWidth); |
| 3054 | break; |
| 3055 | } |
| 3056 | case ISD::ANY_EXTEND: { |
| 3057 | Known = computeKnownBits(Op.getOperand(0), Depth+1); |
| 3058 | Known = Known.zext(BitWidth, false /* ExtendedBitsAreKnownZero */); |
| 3059 | break; |
| 3060 | } |
| 3061 | case ISD::TRUNCATE: { |
| 3062 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3063 | Known = Known.trunc(BitWidth); |
| 3064 | break; |
| 3065 | } |
| 3066 | case ISD::AssertZext: { |
| 3067 | EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); |
| 3068 | APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits()); |
| 3069 | Known = computeKnownBits(Op.getOperand(0), Depth+1); |
| 3070 | Known.Zero |= (~InMask); |
| 3071 | Known.One &= (~Known.Zero); |
| 3072 | break; |
| 3073 | } |
| 3074 | case ISD::FGETSIGN: |
| 3075 | // All bits are zero except the low bit. |
| 3076 | Known.Zero.setBitsFrom(1); |
| 3077 | break; |
| 3078 | case ISD::USUBO: |
| 3079 | case ISD::SSUBO: |
| 3080 | if (Op.getResNo() == 1) { |
| 3081 | // If we know the result of a setcc has the top bits zero, use this info. |
| 3082 | if (TLI->getBooleanContents(Op.getOperand(0).getValueType()) == |
| 3083 | TargetLowering::ZeroOrOneBooleanContent && |
| 3084 | BitWidth > 1) |
| 3085 | Known.Zero.setBitsFrom(1); |
| 3086 | break; |
| 3087 | } |
| 3088 | LLVM_FALLTHROUGH; |
| 3089 | case ISD::SUB: |
| 3090 | case ISD::SUBC: { |
| 3091 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3092 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 3093 | Known = KnownBits::computeForAddSub(/* Add */ false, /* NSW */ false, |
| 3094 | Known, Known2); |
| 3095 | break; |
| 3096 | } |
| 3097 | case ISD::UADDO: |
| 3098 | case ISD::SADDO: |
| 3099 | case ISD::ADDCARRY: |
| 3100 | if (Op.getResNo() == 1) { |
| 3101 | // If we know the result of a setcc has the top bits zero, use this info. |
| 3102 | if (TLI->getBooleanContents(Op.getOperand(0).getValueType()) == |
| 3103 | TargetLowering::ZeroOrOneBooleanContent && |
| 3104 | BitWidth > 1) |
| 3105 | Known.Zero.setBitsFrom(1); |
| 3106 | break; |
| 3107 | } |
| 3108 | LLVM_FALLTHROUGH; |
| 3109 | case ISD::ADD: |
| 3110 | case ISD::ADDC: |
| 3111 | case ISD::ADDE: { |
| 3112 | assert(Op.getResNo() == 0 && "We only compute knownbits for the sum here." ); |
| 3113 | |
| 3114 | // With ADDE and ADDCARRY, a carry bit may be added in. |
| 3115 | KnownBits Carry(1); |
| 3116 | if (Opcode == ISD::ADDE) |
| 3117 | // Can't track carry from glue, set carry to unknown. |
| 3118 | Carry.resetAll(); |
| 3119 | else if (Opcode == ISD::ADDCARRY) |
| 3120 | // TODO: Compute known bits for the carry operand. Not sure if it is worth |
| 3121 | // the trouble (how often will we find a known carry bit). And I haven't |
| 3122 | // tested this very much yet, but something like this might work: |
| 3123 | // Carry = computeKnownBits(Op.getOperand(2), DemandedElts, Depth + 1); |
| 3124 | // Carry = Carry.zextOrTrunc(1, false); |
| 3125 | Carry.resetAll(); |
| 3126 | else |
| 3127 | Carry.setAllZero(); |
| 3128 | |
| 3129 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3130 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 3131 | Known = KnownBits::computeForAddCarry(Known, Known2, Carry); |
| 3132 | break; |
| 3133 | } |
| 3134 | case ISD::SREM: |
| 3135 | if (ConstantSDNode *Rem = isConstOrConstSplat(Op.getOperand(1))) { |
| 3136 | const APInt &RA = Rem->getAPIntValue().abs(); |
| 3137 | if (RA.isPowerOf2()) { |
| 3138 | APInt LowBits = RA - 1; |
| 3139 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3140 | |
| 3141 | // The low bits of the first operand are unchanged by the srem. |
| 3142 | Known.Zero = Known2.Zero & LowBits; |
| 3143 | Known.One = Known2.One & LowBits; |
| 3144 | |
| 3145 | // If the first operand is non-negative or has all low bits zero, then |
| 3146 | // the upper bits are all zero. |
| 3147 | if (Known2.Zero[BitWidth-1] || ((Known2.Zero & LowBits) == LowBits)) |
| 3148 | Known.Zero |= ~LowBits; |
| 3149 | |
| 3150 | // If the first operand is negative and not all low bits are zero, then |
| 3151 | // the upper bits are all one. |
| 3152 | if (Known2.One[BitWidth-1] && ((Known2.One & LowBits) != 0)) |
| 3153 | Known.One |= ~LowBits; |
| 3154 | assert((Known.Zero & Known.One) == 0&&"Bits known to be one AND zero?" ); |
| 3155 | } |
| 3156 | } |
| 3157 | break; |
| 3158 | case ISD::UREM: { |
| 3159 | if (ConstantSDNode *Rem = isConstOrConstSplat(Op.getOperand(1))) { |
| 3160 | const APInt &RA = Rem->getAPIntValue(); |
| 3161 | if (RA.isPowerOf2()) { |
| 3162 | APInt LowBits = (RA - 1); |
| 3163 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3164 | |
| 3165 | // The upper bits are all zero, the lower ones are unchanged. |
| 3166 | Known.Zero = Known2.Zero | ~LowBits; |
| 3167 | Known.One = Known2.One & LowBits; |
| 3168 | break; |
| 3169 | } |
| 3170 | } |
| 3171 | |
| 3172 | // Since the result is less than or equal to either operand, any leading |
| 3173 | // zero bits in either operand must also exist in the result. |
| 3174 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3175 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 3176 | |
| 3177 | uint32_t Leaders = |
| 3178 | std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros()); |
| 3179 | Known.resetAll(); |
| 3180 | Known.Zero.setHighBits(Leaders); |
| 3181 | break; |
| 3182 | } |
| 3183 | case ISD::EXTRACT_ELEMENT: { |
| 3184 | Known = computeKnownBits(Op.getOperand(0), Depth+1); |
| 3185 | const unsigned Index = Op.getConstantOperandVal(1); |
| 3186 | const unsigned EltBitWidth = Op.getValueSizeInBits(); |
| 3187 | |
| 3188 | // Remove low part of known bits mask |
| 3189 | Known.Zero = Known.Zero.getHiBits(Known.getBitWidth() - Index * EltBitWidth); |
| 3190 | Known.One = Known.One.getHiBits(Known.getBitWidth() - Index * EltBitWidth); |
| 3191 | |
| 3192 | // Remove high part of known bit mask |
| 3193 | Known = Known.trunc(EltBitWidth); |
| 3194 | break; |
| 3195 | } |
| 3196 | case ISD::EXTRACT_VECTOR_ELT: { |
| 3197 | SDValue InVec = Op.getOperand(0); |
| 3198 | SDValue EltNo = Op.getOperand(1); |
| 3199 | EVT VecVT = InVec.getValueType(); |
| 3200 | const unsigned EltBitWidth = VecVT.getScalarSizeInBits(); |
| 3201 | const unsigned NumSrcElts = VecVT.getVectorNumElements(); |
| 3202 | // If BitWidth > EltBitWidth the value is anyext:ed. So we do not know |
| 3203 | // anything about the extended bits. |
| 3204 | if (BitWidth > EltBitWidth) |
| 3205 | Known = Known.trunc(EltBitWidth); |
| 3206 | ConstantSDNode *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo); |
| 3207 | if (ConstEltNo && ConstEltNo->getAPIntValue().ult(NumSrcElts)) { |
| 3208 | // If we know the element index, just demand that vector element. |
| 3209 | unsigned Idx = ConstEltNo->getZExtValue(); |
| 3210 | APInt DemandedElt = APInt::getOneBitSet(NumSrcElts, Idx); |
| 3211 | Known = computeKnownBits(InVec, DemandedElt, Depth + 1); |
| 3212 | } else { |
| 3213 | // Unknown element index, so ignore DemandedElts and demand them all. |
| 3214 | Known = computeKnownBits(InVec, Depth + 1); |
| 3215 | } |
| 3216 | if (BitWidth > EltBitWidth) |
| 3217 | Known = Known.zext(BitWidth, false /* => any extend */); |
| 3218 | break; |
| 3219 | } |
| 3220 | case ISD::INSERT_VECTOR_ELT: { |
| 3221 | SDValue InVec = Op.getOperand(0); |
| 3222 | SDValue InVal = Op.getOperand(1); |
| 3223 | SDValue EltNo = Op.getOperand(2); |
| 3224 | |
| 3225 | ConstantSDNode *CEltNo = dyn_cast<ConstantSDNode>(EltNo); |
| 3226 | if (CEltNo && CEltNo->getAPIntValue().ult(NumElts)) { |
| 3227 | // If we know the element index, split the demand between the |
| 3228 | // source vector and the inserted element. |
| 3229 | Known.Zero = Known.One = APInt::getAllOnesValue(BitWidth); |
| 3230 | unsigned EltIdx = CEltNo->getZExtValue(); |
| 3231 | |
| 3232 | // If we demand the inserted element then add its common known bits. |
| 3233 | if (DemandedElts[EltIdx]) { |
| 3234 | Known2 = computeKnownBits(InVal, Depth + 1); |
| 3235 | Known.One &= Known2.One.zextOrTrunc(Known.One.getBitWidth()); |
| 3236 | Known.Zero &= Known2.Zero.zextOrTrunc(Known.Zero.getBitWidth()); |
| 3237 | } |
| 3238 | |
| 3239 | // If we demand the source vector then add its common known bits, ensuring |
| 3240 | // that we don't demand the inserted element. |
| 3241 | APInt VectorElts = DemandedElts & ~(APInt::getOneBitSet(NumElts, EltIdx)); |
| 3242 | if (!!VectorElts) { |
| 3243 | Known2 = computeKnownBits(InVec, VectorElts, Depth + 1); |
| 3244 | Known.One &= Known2.One; |
| 3245 | Known.Zero &= Known2.Zero; |
| 3246 | } |
| 3247 | } else { |
| 3248 | // Unknown element index, so ignore DemandedElts and demand them all. |
| 3249 | Known = computeKnownBits(InVec, Depth + 1); |
| 3250 | Known2 = computeKnownBits(InVal, Depth + 1); |
| 3251 | Known.One &= Known2.One.zextOrTrunc(Known.One.getBitWidth()); |
| 3252 | Known.Zero &= Known2.Zero.zextOrTrunc(Known.Zero.getBitWidth()); |
| 3253 | } |
| 3254 | break; |
| 3255 | } |
| 3256 | case ISD::BITREVERSE: { |
| 3257 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3258 | Known.Zero = Known2.Zero.reverseBits(); |
| 3259 | Known.One = Known2.One.reverseBits(); |
| 3260 | break; |
| 3261 | } |
| 3262 | case ISD::BSWAP: { |
| 3263 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3264 | Known.Zero = Known2.Zero.byteSwap(); |
| 3265 | Known.One = Known2.One.byteSwap(); |
| 3266 | break; |
| 3267 | } |
| 3268 | case ISD::ABS: { |
| 3269 | Known2 = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3270 | |
| 3271 | // If the source's MSB is zero then we know the rest of the bits already. |
| 3272 | if (Known2.isNonNegative()) { |
| 3273 | Known.Zero = Known2.Zero; |
| 3274 | Known.One = Known2.One; |
| 3275 | break; |
| 3276 | } |
| 3277 | |
| 3278 | // We only know that the absolute values's MSB will be zero iff there is |
| 3279 | // a set bit that isn't the sign bit (otherwise it could be INT_MIN). |
| 3280 | Known2.One.clearSignBit(); |
| 3281 | if (Known2.One.getBoolValue()) { |
| 3282 | Known.Zero = APInt::getSignMask(BitWidth); |
| 3283 | break; |
| 3284 | } |
| 3285 | break; |
| 3286 | } |
| 3287 | case ISD::UMIN: { |
| 3288 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3289 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 3290 | |
| 3291 | // UMIN - we know that the result will have the maximum of the |
| 3292 | // known zero leading bits of the inputs. |
| 3293 | unsigned LeadZero = Known.countMinLeadingZeros(); |
| 3294 | LeadZero = std::max(LeadZero, Known2.countMinLeadingZeros()); |
| 3295 | |
| 3296 | Known.Zero &= Known2.Zero; |
| 3297 | Known.One &= Known2.One; |
| 3298 | Known.Zero.setHighBits(LeadZero); |
| 3299 | break; |
| 3300 | } |
| 3301 | case ISD::UMAX: { |
| 3302 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3303 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 3304 | |
| 3305 | // UMAX - we know that the result will have the maximum of the |
| 3306 | // known one leading bits of the inputs. |
| 3307 | unsigned LeadOne = Known.countMinLeadingOnes(); |
| 3308 | LeadOne = std::max(LeadOne, Known2.countMinLeadingOnes()); |
| 3309 | |
| 3310 | Known.Zero &= Known2.Zero; |
| 3311 | Known.One &= Known2.One; |
| 3312 | Known.One.setHighBits(LeadOne); |
| 3313 | break; |
| 3314 | } |
| 3315 | case ISD::SMIN: |
| 3316 | case ISD::SMAX: { |
| 3317 | // If we have a clamp pattern, we know that the number of sign bits will be |
| 3318 | // the minimum of the clamp min/max range. |
| 3319 | bool IsMax = (Opcode == ISD::SMAX); |
| 3320 | ConstantSDNode *CstLow = nullptr, *CstHigh = nullptr; |
| 3321 | if ((CstLow = isConstOrConstSplat(Op.getOperand(1), DemandedElts))) |
| 3322 | if (Op.getOperand(0).getOpcode() == (IsMax ? ISD::SMIN : ISD::SMAX)) |
| 3323 | CstHigh = |
| 3324 | isConstOrConstSplat(Op.getOperand(0).getOperand(1), DemandedElts); |
| 3325 | if (CstLow && CstHigh) { |
| 3326 | if (!IsMax) |
| 3327 | std::swap(CstLow, CstHigh); |
| 3328 | |
| 3329 | const APInt &ValueLow = CstLow->getAPIntValue(); |
| 3330 | const APInt &ValueHigh = CstHigh->getAPIntValue(); |
| 3331 | if (ValueLow.sle(ValueHigh)) { |
| 3332 | unsigned LowSignBits = ValueLow.getNumSignBits(); |
| 3333 | unsigned HighSignBits = ValueHigh.getNumSignBits(); |
| 3334 | unsigned MinSignBits = std::min(LowSignBits, HighSignBits); |
| 3335 | if (ValueLow.isNegative() && ValueHigh.isNegative()) { |
| 3336 | Known.One.setHighBits(MinSignBits); |
| 3337 | break; |
| 3338 | } |
| 3339 | if (ValueLow.isNonNegative() && ValueHigh.isNonNegative()) { |
| 3340 | Known.Zero.setHighBits(MinSignBits); |
| 3341 | break; |
| 3342 | } |
| 3343 | } |
| 3344 | } |
| 3345 | |
| 3346 | // Fallback - just get the shared known bits of the operands. |
| 3347 | Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); |
| 3348 | if (Known.isUnknown()) break; // Early-out |
| 3349 | Known2 = computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); |
| 3350 | Known.Zero &= Known2.Zero; |
| 3351 | Known.One &= Known2.One; |
| 3352 | break; |
| 3353 | } |
| 3354 | case ISD::FrameIndex: |
| 3355 | case ISD::TargetFrameIndex: |
| 3356 | TLI->computeKnownBitsForFrameIndex(Op, Known, DemandedElts, *this, Depth); |
| 3357 | break; |
| 3358 | |
| 3359 | default: |
| 3360 | if (Opcode < ISD::BUILTIN_OP_END) |
| 3361 | break; |
| 3362 | LLVM_FALLTHROUGH; |
| 3363 | case ISD::INTRINSIC_WO_CHAIN: |
| 3364 | case ISD::INTRINSIC_W_CHAIN: |
| 3365 | case ISD::INTRINSIC_VOID: |
| 3366 | // Allow the target to implement this method for its nodes. |
| 3367 | TLI->computeKnownBitsForTargetNode(Op, Known, DemandedElts, *this, Depth); |
| 3368 | break; |
| 3369 | } |
| 3370 | |
| 3371 | assert(!Known.hasConflict() && "Bits known to be one AND zero?" ); |
| 3372 | return Known; |
| 3373 | } |
| 3374 | |
| 3375 | SelectionDAG::OverflowKind SelectionDAG::computeOverflowKind(SDValue N0, |
| 3376 | SDValue N1) const { |
| 3377 | // X + 0 never overflow |
| 3378 | if (isNullConstant(N1)) |
| 3379 | return OFK_Never; |
| 3380 | |
| 3381 | KnownBits N1Known = computeKnownBits(N1); |
| 3382 | if (N1Known.Zero.getBoolValue()) { |
| 3383 | KnownBits N0Known = computeKnownBits(N0); |
| 3384 | |
| 3385 | bool overflow; |
| 3386 | (void)(~N0Known.Zero).uadd_ov(~N1Known.Zero, overflow); |
| 3387 | if (!overflow) |
| 3388 | return OFK_Never; |
| 3389 | } |
| 3390 | |
| 3391 | // mulhi + 1 never overflow |
| 3392 | if (N0.getOpcode() == ISD::UMUL_LOHI && N0.getResNo() == 1 && |
| 3393 | (~N1Known.Zero & 0x01) == ~N1Known.Zero) |
| 3394 | return OFK_Never; |
| 3395 | |
| 3396 | if (N1.getOpcode() == ISD::UMUL_LOHI && N1.getResNo() == 1) { |
| 3397 | KnownBits N0Known = computeKnownBits(N0); |
| 3398 | |
| 3399 | if ((~N0Known.Zero & 0x01) == ~N0Known.Zero) |
| 3400 | return OFK_Never; |
| 3401 | } |
| 3402 | |
| 3403 | return OFK_Sometime; |
| 3404 | } |
| 3405 | |
| 3406 | bool SelectionDAG::isKnownToBeAPowerOfTwo(SDValue Val) const { |
| 3407 | EVT OpVT = Val.getValueType(); |
| 3408 | unsigned BitWidth = OpVT.getScalarSizeInBits(); |
| 3409 | |
| 3410 | // Is the constant a known power of 2? |
| 3411 | if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Val)) |
| 3412 | return Const->getAPIntValue().zextOrTrunc(BitWidth).isPowerOf2(); |
| 3413 | |
| 3414 | // A left-shift of a constant one will have exactly one bit set because |
| 3415 | // shifting the bit off the end is undefined. |
| 3416 | if (Val.getOpcode() == ISD::SHL) { |
| 3417 | auto *C = isConstOrConstSplat(Val.getOperand(0)); |
| 3418 | if (C && C->getAPIntValue() == 1) |
| 3419 | return true; |
| 3420 | } |
| 3421 | |
| 3422 | // Similarly, a logical right-shift of a constant sign-bit will have exactly |
| 3423 | // one bit set. |
| 3424 | if (Val.getOpcode() == ISD::SRL) { |
| 3425 | auto *C = isConstOrConstSplat(Val.getOperand(0)); |
| 3426 | if (C && C->getAPIntValue().isSignMask()) |
| 3427 | return true; |
| 3428 | } |
| 3429 | |
| 3430 | // Are all operands of a build vector constant powers of two? |
| 3431 | if (Val.getOpcode() == ISD::BUILD_VECTOR) |
| 3432 | if (llvm::all_of(Val->ops(), [BitWidth](SDValue E) { |
| 3433 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(E)) |
| 3434 | return C->getAPIntValue().zextOrTrunc(BitWidth).isPowerOf2(); |
| 3435 | return false; |
| 3436 | })) |
| 3437 | return true; |
| 3438 | |
| 3439 | // More could be done here, though the above checks are enough |
| 3440 | // to handle some common cases. |
| 3441 | |
| 3442 | // Fall back to computeKnownBits to catch other known cases. |
| 3443 | KnownBits Known = computeKnownBits(Val); |
| 3444 | return (Known.countMaxPopulation() == 1) && (Known.countMinPopulation() == 1); |
| 3445 | } |
| 3446 | |
| 3447 | unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const { |
| 3448 | EVT VT = Op.getValueType(); |
| 3449 | APInt DemandedElts = VT.isVector() |
| 3450 | ? APInt::getAllOnesValue(VT.getVectorNumElements()) |
| 3451 | : APInt(1, 1); |
| 3452 | return ComputeNumSignBits(Op, DemandedElts, Depth); |
| 3453 | } |
| 3454 | |
| 3455 | unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, const APInt &DemandedElts, |
| 3456 | unsigned Depth) const { |
| 3457 | EVT VT = Op.getValueType(); |
| 3458 | assert((VT.isInteger() || VT.isFloatingPoint()) && "Invalid VT!" ); |
| 3459 | unsigned VTBits = VT.getScalarSizeInBits(); |
| 3460 | unsigned NumElts = DemandedElts.getBitWidth(); |
| 3461 | unsigned Tmp, Tmp2; |
| 3462 | unsigned FirstAnswer = 1; |
| 3463 | |
| 3464 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { |
| 3465 | const APInt &Val = C->getAPIntValue(); |
| 3466 | return Val.getNumSignBits(); |
| 3467 | } |
| 3468 | |
| 3469 | if (Depth == 6) |
| 3470 | return 1; // Limit search depth. |
| 3471 | |
| 3472 | if (!DemandedElts) |
| 3473 | return 1; // No demanded elts, better to assume we don't know anything. |
| 3474 | |
| 3475 | unsigned Opcode = Op.getOpcode(); |
| 3476 | switch (Opcode) { |
| 3477 | default: break; |
| 3478 | case ISD::AssertSext: |
| 3479 | Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); |
| 3480 | return VTBits-Tmp+1; |
| 3481 | case ISD::AssertZext: |
| 3482 | Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); |
| 3483 | return VTBits-Tmp; |
| 3484 | |
| 3485 | case ISD::BUILD_VECTOR: |
| 3486 | Tmp = VTBits; |
| 3487 | for (unsigned i = 0, e = Op.getNumOperands(); (i < e) && (Tmp > 1); ++i) { |
| 3488 | if (!DemandedElts[i]) |
| 3489 | continue; |
| 3490 | |
| 3491 | SDValue SrcOp = Op.getOperand(i); |
| 3492 | Tmp2 = ComputeNumSignBits(Op.getOperand(i), Depth + 1); |
| 3493 | |
| 3494 | // BUILD_VECTOR can implicitly truncate sources, we must handle this. |
| 3495 | if (SrcOp.getValueSizeInBits() != VTBits) { |
| 3496 | assert(SrcOp.getValueSizeInBits() > VTBits && |
| 3497 | "Expected BUILD_VECTOR implicit truncation" ); |
| 3498 | unsigned = SrcOp.getValueSizeInBits() - VTBits; |
| 3499 | Tmp2 = (Tmp2 > ExtraBits ? Tmp2 - ExtraBits : 1); |
| 3500 | } |
| 3501 | Tmp = std::min(Tmp, Tmp2); |
| 3502 | } |
| 3503 | return Tmp; |
| 3504 | |
| 3505 | case ISD::VECTOR_SHUFFLE: { |
| 3506 | // Collect the minimum number of sign bits that are shared by every vector |
| 3507 | // element referenced by the shuffle. |
| 3508 | APInt DemandedLHS(NumElts, 0), DemandedRHS(NumElts, 0); |
| 3509 | const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op); |
| 3510 | assert(NumElts == SVN->getMask().size() && "Unexpected vector size" ); |
| 3511 | for (unsigned i = 0; i != NumElts; ++i) { |
| 3512 | int M = SVN->getMaskElt(i); |
| 3513 | if (!DemandedElts[i]) |
| 3514 | continue; |
| 3515 | // For UNDEF elements, we don't know anything about the common state of |
| 3516 | // the shuffle result. |
| 3517 | if (M < 0) |
| 3518 | return 1; |
| 3519 | if ((unsigned)M < NumElts) |
| 3520 | DemandedLHS.setBit((unsigned)M % NumElts); |
| 3521 | else |
| 3522 | DemandedRHS.setBit((unsigned)M % NumElts); |
| 3523 | } |
| 3524 | Tmp = std::numeric_limits<unsigned>::max(); |
| 3525 | if (!!DemandedLHS) |
| 3526 | Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedLHS, Depth + 1); |
| 3527 | if (!!DemandedRHS) { |
| 3528 | Tmp2 = ComputeNumSignBits(Op.getOperand(1), DemandedRHS, Depth + 1); |
| 3529 | Tmp = std::min(Tmp, Tmp2); |
| 3530 | } |
| 3531 | // If we don't know anything, early out and try computeKnownBits fall-back. |
| 3532 | if (Tmp == 1) |
| 3533 | break; |
| 3534 | assert(Tmp <= VTBits && "Failed to determine minimum sign bits" ); |
| 3535 | return Tmp; |
| 3536 | } |
| 3537 | |
| 3538 | case ISD::BITCAST: { |
| 3539 | SDValue N0 = Op.getOperand(0); |
| 3540 | EVT SrcVT = N0.getValueType(); |
| 3541 | unsigned SrcBits = SrcVT.getScalarSizeInBits(); |
| 3542 | |
| 3543 | // Ignore bitcasts from unsupported types.. |
| 3544 | if (!(SrcVT.isInteger() || SrcVT.isFloatingPoint())) |
| 3545 | break; |
| 3546 | |
| 3547 | // Fast handling of 'identity' bitcasts. |
| 3548 | if (VTBits == SrcBits) |
| 3549 | return ComputeNumSignBits(N0, DemandedElts, Depth + 1); |
| 3550 | |
| 3551 | bool IsLE = getDataLayout().isLittleEndian(); |
| 3552 | |
| 3553 | // Bitcast 'large element' scalar/vector to 'small element' vector. |
| 3554 | if ((SrcBits % VTBits) == 0) { |
| 3555 | assert(VT.isVector() && "Expected bitcast to vector" ); |
| 3556 | |
| 3557 | unsigned Scale = SrcBits / VTBits; |
| 3558 | APInt SrcDemandedElts(NumElts / Scale, 0); |
| 3559 | for (unsigned i = 0; i != NumElts; ++i) |
| 3560 | if (DemandedElts[i]) |
| 3561 | SrcDemandedElts.setBit(i / Scale); |
| 3562 | |
| 3563 | // Fast case - sign splat can be simply split across the small elements. |
| 3564 | Tmp = ComputeNumSignBits(N0, SrcDemandedElts, Depth + 1); |
| 3565 | if (Tmp == SrcBits) |
| 3566 | return VTBits; |
| 3567 | |
| 3568 | // Slow case - determine how far the sign extends into each sub-element. |
| 3569 | Tmp2 = VTBits; |
| 3570 | for (unsigned i = 0; i != NumElts; ++i) |
| 3571 | if (DemandedElts[i]) { |
| 3572 | unsigned SubOffset = i % Scale; |
| 3573 | SubOffset = (IsLE ? ((Scale - 1) - SubOffset) : SubOffset); |
| 3574 | SubOffset = SubOffset * VTBits; |
| 3575 | if (Tmp <= SubOffset) |
| 3576 | return 1; |
| 3577 | Tmp2 = std::min(Tmp2, Tmp - SubOffset); |
| 3578 | } |
| 3579 | return Tmp2; |
| 3580 | } |
| 3581 | break; |
| 3582 | } |
| 3583 | |
| 3584 | case ISD::SIGN_EXTEND: |
| 3585 | Tmp = VTBits - Op.getOperand(0).getScalarValueSizeInBits(); |
| 3586 | return ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth+1) + Tmp; |
| 3587 | case ISD::SIGN_EXTEND_INREG: |
| 3588 | // Max of the input and what this extends. |
| 3589 | Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarSizeInBits(); |
| 3590 | Tmp = VTBits-Tmp+1; |
| 3591 | Tmp2 = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth+1); |
| 3592 | return std::max(Tmp, Tmp2); |
| 3593 | case ISD::SIGN_EXTEND_VECTOR_INREG: { |
| 3594 | SDValue Src = Op.getOperand(0); |
| 3595 | EVT SrcVT = Src.getValueType(); |
| 3596 | APInt DemandedSrcElts = DemandedElts.zextOrSelf(SrcVT.getVectorNumElements()); |
| 3597 | Tmp = VTBits - SrcVT.getScalarSizeInBits(); |
| 3598 | return ComputeNumSignBits(Src, DemandedSrcElts, Depth+1) + Tmp; |
| 3599 | } |
| 3600 | |
| 3601 | case ISD::SRA: |
| 3602 | Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth+1); |
| 3603 | // SRA X, C -> adds C sign bits. |
| 3604 | if (ConstantSDNode *C = |
| 3605 | isConstOrConstSplat(Op.getOperand(1), DemandedElts)) { |
| 3606 | APInt ShiftVal = C->getAPIntValue(); |
| 3607 | ShiftVal += Tmp; |
| 3608 | Tmp = ShiftVal.uge(VTBits) ? VTBits : ShiftVal.getZExtValue(); |
| 3609 | } |
| 3610 | return Tmp; |
| 3611 | case ISD::SHL: |
| 3612 | if (ConstantSDNode *C = |
| 3613 | isConstOrConstSplat(Op.getOperand(1), DemandedElts)) { |
| 3614 | // shl destroys sign bits. |
| 3615 | Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth+1); |
| 3616 | if (C->getAPIntValue().uge(VTBits) || // Bad shift. |
| 3617 | C->getAPIntValue().uge(Tmp)) break; // Shifted all sign bits out. |
| 3618 | return Tmp - C->getZExtValue(); |
| 3619 | } |
| 3620 | break; |
| 3621 | case ISD::AND: |
| 3622 | case ISD::OR: |
| 3623 | case ISD::XOR: // NOT is handled here. |
| 3624 | // Logical binary ops preserve the number of sign bits at the worst. |
| 3625 | Tmp = ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth+1); |
| 3626 | if (Tmp != 1) { |
| 3627 | Tmp2 = ComputeNumSignBits(Op.getOperand(1), DemandedElts, Depth+1); |
| 3628 | FirstAnswer = std::min(Tmp, Tmp2); |
| 3629 | // We computed what we know about the sign bits as our first |
| 3630 | // answer. Now proceed to the generic code that uses |
| 3631 | // computeKnownBits, and pick whichever answer is better. |
| 3632 | } |
| 3633 | break; |
| 3634 | |
| 3635 | case ISD::SELECT: |
| 3636 | case ISD::VSELECT: |
| 3637 | Tmp = ComputeNumSignBits(Op.getOperand(1), DemandedElts, Depth+1); |
| 3638 | if (Tmp == 1) return 1; // Early out. |
| 3639 | Tmp2 = ComputeNumSignBits(Op.getOperand(2), DemandedElts, Depth+1); |
| 3640 | return std::min(Tmp, Tmp2); |
| 3641 | case ISD::SELECT_CC: |
| 3642 | Tmp = ComputeNumSignBits(Op.getOperand(2), DemandedElts, Depth+1); |
| 3643 | if (Tmp == 1) return 1; // Early out. |
| 3644 | Tmp2 = ComputeNumSignBits(Op.getOperand(3), DemandedElts, Depth+1); |
| 3645 | return std::min(Tmp, Tmp2); |
| 3646 | |
| 3647 | case ISD::SMIN: |
| 3648 | case ISD::SMAX: { |
| 3649 | // If we have a clamp pattern, we know that the number of sign bits will be |
| 3650 | // the minimum of the clamp min/max range. |
| 3651 | bool IsMax = (Opcode == ISD::SMAX); |
| 3652 | ConstantSDNode *CstLow = nullptr, *CstHigh = nullptr; |
| 3653 | if ((CstLow = isConstOrConstSplat(Op.getOperand(1), DemandedElts))) |
| 3654 | if (Op.getOperand(0).getOpcode() == (IsMax ? ISD::SMIN : ISD::SMAX)) |
| 3655 | CstHigh = |
| 3656 | isConstOrConstSplat(Op.getOperand(0).getOperand(1), DemandedElts); |
| 3657 | if (CstLow && CstHigh) { |
| 3658 | if (!IsMax) |
| 3659 | std::swap(CstLow, CstHigh); |
| 3660 | if (CstLow->getAPIntValue().sle(CstHigh->getAPIntValue())) { |
| 3661 | Tmp = CstLow->getAPIntValue().getNumSignBits(); |
| 3662 | Tmp2 = CstHigh->getAPIntValue().getNumSignBits(); |
| 3663 | return std::min(Tmp, Tmp2); |
| 3664 | } |
| 3665 | } |
| 3666 | |
| 3667 | // Fallback - just get the minimum number of sign bits of the operands. |
| 3668 | Tmp = ComputeNumSignBits(Op.getOperand(0), Depth + 1); |
| 3669 | if (Tmp == 1) |
| 3670 | return 1; // Early out. |
| 3671 | Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth + 1); |
| 3672 | return std::min(Tmp, Tmp2); |
| 3673 | } |
| 3674 | case ISD::UMIN: |
| 3675 | case ISD::UMAX: |
| 3676 | Tmp = ComputeNumSignBits(Op.getOperand(0), Depth + 1); |
| 3677 | if (Tmp == 1) |
| 3678 | return 1; // Early out. |
| 3679 | Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth + 1); |
| 3680 | return std::min(Tmp, Tmp2); |
| 3681 | case ISD::SADDO: |
| 3682 | case ISD::UADDO: |
| 3683 | case ISD::SSUBO: |
| 3684 | case ISD::USUBO: |
| 3685 | case ISD::SMULO: |
| 3686 | case ISD::UMULO: |
| 3687 | if (Op.getResNo() != 1) |
| 3688 | break; |
| 3689 | // The boolean result conforms to getBooleanContents. Fall through. |
| 3690 | // If setcc returns 0/-1, all bits are sign bits. |
| 3691 | // We know that we have an integer-based boolean since these operations |
| 3692 | // are only available for integer. |
| 3693 | if (TLI->getBooleanContents(VT.isVector(), false) == |
| 3694 | TargetLowering::ZeroOrNegativeOneBooleanContent) |
| 3695 | return VTBits; |
| 3696 | break; |
| 3697 | case ISD::SETCC: |
| 3698 | // If setcc returns 0/-1, all bits are sign bits. |
| 3699 | if (TLI->getBooleanContents(Op.getOperand(0).getValueType()) == |
| 3700 | TargetLowering::ZeroOrNegativeOneBooleanContent) |
| 3701 | return VTBits; |
| 3702 | break; |
| 3703 | case ISD::ROTL: |
| 3704 | case ISD::ROTR: |
| 3705 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { |
| 3706 | unsigned RotAmt = C->getAPIntValue().urem(VTBits); |
| 3707 | |
| 3708 | // Handle rotate right by N like a rotate left by 32-N. |
| 3709 | if (Opcode == ISD::ROTR) |
| 3710 | RotAmt = (VTBits - RotAmt) % VTBits; |
| 3711 | |
| 3712 | // If we aren't rotating out all of the known-in sign bits, return the |
| 3713 | // number that are left. This handles rotl(sext(x), 1) for example. |
| 3714 | Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); |
| 3715 | if (Tmp > (RotAmt + 1)) return (Tmp - RotAmt); |
| 3716 | } |
| 3717 | break; |
| 3718 | case ISD::ADD: |
| 3719 | case ISD::ADDC: |
| 3720 | // Add can have at most one carry bit. Thus we know that the output |
| 3721 | // is, at worst, one more bit than the inputs. |
| 3722 | Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); |
| 3723 | if (Tmp == 1) return 1; // Early out. |
| 3724 | |
| 3725 | // Special case decrementing a value (ADD X, -1): |
| 3726 | if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1))) |
| 3727 | if (CRHS->isAllOnesValue()) { |
| 3728 | KnownBits Known = computeKnownBits(Op.getOperand(0), Depth+1); |
| 3729 | |
| 3730 | // If the input is known to be 0 or 1, the output is 0/-1, which is all |
| 3731 | // sign bits set. |
| 3732 | if ((Known.Zero | 1).isAllOnesValue()) |
| 3733 | return VTBits; |
| 3734 | |
| 3735 | // If we are subtracting one from a positive number, there is no carry |
| 3736 | // out of the result. |
| 3737 | if (Known.isNonNegative()) |
| 3738 | return Tmp; |
| 3739 | } |
| 3740 | |
| 3741 | Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); |
| 3742 | if (Tmp2 == 1) return 1; |
| 3743 | return std::min(Tmp, Tmp2)-1; |
| 3744 | |
| 3745 | case ISD::SUB: |
| 3746 | Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); |
| 3747 | if (Tmp2 == 1) return 1; |
| 3748 | |
| 3749 | // Handle NEG. |
| 3750 | if (ConstantSDNode *CLHS = isConstOrConstSplat(Op.getOperand(0))) |
| 3751 | if (CLHS->isNullValue()) { |
| 3752 | KnownBits Known = computeKnownBits(Op.getOperand(1), Depth+1); |
| 3753 | // If the input is known to be 0 or 1, the output is 0/-1, which is all |
| 3754 | // sign bits set. |
| 3755 | if ((Known.Zero | 1).isAllOnesValue()) |
| 3756 | return VTBits; |
| 3757 | |
| 3758 | // If the input is known to be positive (the sign bit is known clear), |
| 3759 | // the output of the NEG has the same number of sign bits as the input. |
| 3760 | if (Known.isNonNegative()) |
| 3761 | return Tmp2; |
| 3762 | |
| 3763 | // Otherwise, we treat this like a SUB. |
| 3764 | } |
| 3765 | |
| 3766 | // Sub can have at most one carry bit. Thus we know that the output |
| 3767 | // is, at worst, one more bit than the inputs. |
| 3768 | Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); |
| 3769 | if (Tmp == 1) return 1; // Early out. |
| 3770 | return std::min(Tmp, Tmp2)-1; |
| 3771 | case ISD::TRUNCATE: { |
| 3772 | // Check if the sign bits of source go down as far as the truncated value. |
| 3773 | unsigned NumSrcBits = Op.getOperand(0).getScalarValueSizeInBits(); |
| 3774 | unsigned NumSrcSignBits = ComputeNumSignBits(Op.getOperand(0), Depth + 1); |
| 3775 | if (NumSrcSignBits > (NumSrcBits - VTBits)) |
| 3776 | return NumSrcSignBits - (NumSrcBits - VTBits); |
| 3777 | break; |
| 3778 | } |
| 3779 | case ISD::EXTRACT_ELEMENT: { |
| 3780 | const int KnownSign = ComputeNumSignBits(Op.getOperand(0), Depth+1); |
| 3781 | const int BitWidth = Op.getValueSizeInBits(); |
| 3782 | const int Items = Op.getOperand(0).getValueSizeInBits() / BitWidth; |
| 3783 | |
| 3784 | // Get reverse index (starting from 1), Op1 value indexes elements from |
| 3785 | // little end. Sign starts at big end. |
| 3786 | const int rIndex = Items - 1 - Op.getConstantOperandVal(1); |
| 3787 | |
| 3788 | // If the sign portion ends in our element the subtraction gives correct |
| 3789 | // result. Otherwise it gives either negative or > bitwidth result |
| 3790 | return std::max(std::min(KnownSign - rIndex * BitWidth, BitWidth), 0); |
| 3791 | } |
| 3792 | case ISD::INSERT_VECTOR_ELT: { |
| 3793 | SDValue InVec = Op.getOperand(0); |
| 3794 | SDValue InVal = Op.getOperand(1); |
| 3795 | SDValue EltNo = Op.getOperand(2); |
| 3796 | |
| 3797 | ConstantSDNode *CEltNo = dyn_cast<ConstantSDNode>(EltNo); |
| 3798 | if (CEltNo && CEltNo->getAPIntValue().ult(NumElts)) { |
| 3799 | // If we know the element index, split the demand between the |
| 3800 | // source vector and the inserted element. |
| 3801 | unsigned EltIdx = CEltNo->getZExtValue(); |
| 3802 | |
| 3803 | // If we demand the inserted element then get its sign bits. |
| 3804 | Tmp = std::numeric_limits<unsigned>::max(); |
| 3805 | if (DemandedElts[EltIdx]) { |
| 3806 | // TODO - handle implicit truncation of inserted elements. |
| 3807 | if (InVal.getScalarValueSizeInBits() != VTBits) |
| 3808 | break; |
| 3809 | Tmp = ComputeNumSignBits(InVal, Depth + 1); |
| 3810 | } |
| 3811 | |
| 3812 | // If we demand the source vector then get its sign bits, and determine |
| 3813 | // the minimum. |
| 3814 | APInt VectorElts = DemandedElts; |
| 3815 | VectorElts.clearBit(EltIdx); |
| 3816 | if (!!VectorElts) { |
| 3817 | Tmp2 = ComputeNumSignBits(InVec, VectorElts, Depth + 1); |
| 3818 | Tmp = std::min(Tmp, Tmp2); |
| 3819 | } |
| 3820 | } else { |
| 3821 | // Unknown element index, so ignore DemandedElts and demand them all. |
| 3822 | Tmp = ComputeNumSignBits(InVec, Depth + 1); |
| 3823 | Tmp2 = ComputeNumSignBits(InVal, Depth + 1); |
| 3824 | Tmp = std::min(Tmp, Tmp2); |
| 3825 | } |
| 3826 | assert(Tmp <= VTBits && "Failed to determine minimum sign bits" ); |
| 3827 | return Tmp; |
| 3828 | } |
| 3829 | case ISD::EXTRACT_VECTOR_ELT: { |
| 3830 | SDValue InVec = Op.getOperand(0); |
| 3831 | SDValue EltNo = Op.getOperand(1); |
| 3832 | EVT VecVT = InVec.getValueType(); |
| 3833 | const unsigned BitWidth = Op.getValueSizeInBits(); |
| 3834 | const unsigned EltBitWidth = Op.getOperand(0).getScalarValueSizeInBits(); |
| 3835 | const unsigned NumSrcElts = VecVT.getVectorNumElements(); |
| 3836 | |
| 3837 | // If BitWidth > EltBitWidth the value is anyext:ed, and we do not know |
| 3838 | // anything about sign bits. But if the sizes match we can derive knowledge |
| 3839 | // about sign bits from the vector operand. |
| 3840 | if (BitWidth != EltBitWidth) |
| 3841 | break; |
| 3842 | |
| 3843 | // If we know the element index, just demand that vector element, else for |
| 3844 | // an unknown element index, ignore DemandedElts and demand them all. |
| 3845 | APInt DemandedSrcElts = APInt::getAllOnesValue(NumSrcElts); |
| 3846 | ConstantSDNode *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo); |
| 3847 | if (ConstEltNo && ConstEltNo->getAPIntValue().ult(NumSrcElts)) |
| 3848 | DemandedSrcElts = |
| 3849 | APInt::getOneBitSet(NumSrcElts, ConstEltNo->getZExtValue()); |
| 3850 | |
| 3851 | return ComputeNumSignBits(InVec, DemandedSrcElts, Depth + 1); |
| 3852 | } |
| 3853 | case ISD::EXTRACT_SUBVECTOR: { |
| 3854 | // If we know the element index, just demand that subvector elements, |
| 3855 | // otherwise demand them all. |
| 3856 | SDValue Src = Op.getOperand(0); |
| 3857 | ConstantSDNode *SubIdx = dyn_cast<ConstantSDNode>(Op.getOperand(1)); |
| 3858 | unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); |
| 3859 | if (SubIdx && SubIdx->getAPIntValue().ule(NumSrcElts - NumElts)) { |
| 3860 | // Offset the demanded elts by the subvector index. |
| 3861 | uint64_t Idx = SubIdx->getZExtValue(); |
| 3862 | APInt DemandedSrc = DemandedElts.zextOrSelf(NumSrcElts).shl(Idx); |
| 3863 | return ComputeNumSignBits(Src, DemandedSrc, Depth + 1); |
| 3864 | } |
| 3865 | return ComputeNumSignBits(Src, Depth + 1); |
| 3866 | } |
| 3867 | case ISD::CONCAT_VECTORS: { |
| 3868 | // Determine the minimum number of sign bits across all demanded |
| 3869 | // elts of the input vectors. Early out if the result is already 1. |
| 3870 | Tmp = std::numeric_limits<unsigned>::max(); |
| 3871 | EVT SubVectorVT = Op.getOperand(0).getValueType(); |
| 3872 | unsigned NumSubVectorElts = SubVectorVT.getVectorNumElements(); |
| 3873 | unsigned NumSubVectors = Op.getNumOperands(); |
| 3874 | for (unsigned i = 0; (i < NumSubVectors) && (Tmp > 1); ++i) { |
| 3875 | APInt DemandedSub = DemandedElts.lshr(i * NumSubVectorElts); |
| 3876 | DemandedSub = DemandedSub.trunc(NumSubVectorElts); |
| 3877 | if (!DemandedSub) |
| 3878 | continue; |
| 3879 | Tmp2 = ComputeNumSignBits(Op.getOperand(i), DemandedSub, Depth + 1); |
| 3880 | Tmp = std::min(Tmp, Tmp2); |
| 3881 | } |
| 3882 | assert(Tmp <= VTBits && "Failed to determine minimum sign bits" ); |
| 3883 | return Tmp; |
| 3884 | } |
| 3885 | case ISD::INSERT_SUBVECTOR: { |
| 3886 | // If we know the element index, demand any elements from the subvector and |
| 3887 | // the remainder from the src its inserted into, otherwise demand them all. |
| 3888 | SDValue Src = Op.getOperand(0); |
| 3889 | SDValue Sub = Op.getOperand(1); |
| 3890 | auto *SubIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2)); |
| 3891 | unsigned NumSubElts = Sub.getValueType().getVectorNumElements(); |
| 3892 | if (SubIdx && SubIdx->getAPIntValue().ule(NumElts - NumSubElts)) { |
| 3893 | Tmp = std::numeric_limits<unsigned>::max(); |
| 3894 | uint64_t Idx = SubIdx->getZExtValue(); |
| 3895 | APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx); |
| 3896 | if (!!DemandedSubElts) { |
| 3897 | Tmp = ComputeNumSignBits(Sub, DemandedSubElts, Depth + 1); |
| 3898 | if (Tmp == 1) return 1; // early-out |
| 3899 | } |
| 3900 | APInt SubMask = APInt::getBitsSet(NumElts, Idx, Idx + NumSubElts); |
| 3901 | APInt DemandedSrcElts = DemandedElts & ~SubMask; |
| 3902 | if (!!DemandedSrcElts) { |
| 3903 | Tmp2 = ComputeNumSignBits(Src, DemandedSrcElts, Depth + 1); |
| 3904 | Tmp = std::min(Tmp, Tmp2); |
| 3905 | } |
| 3906 | assert(Tmp <= VTBits && "Failed to determine minimum sign bits" ); |
| 3907 | return Tmp; |
| 3908 | } |
| 3909 | |
| 3910 | // Not able to determine the index so just assume worst case. |
| 3911 | Tmp = ComputeNumSignBits(Sub, Depth + 1); |
| 3912 | if (Tmp == 1) return 1; // early-out |
| 3913 | Tmp2 = ComputeNumSignBits(Src, Depth + 1); |
| 3914 | Tmp = std::min(Tmp, Tmp2); |
| 3915 | assert(Tmp <= VTBits && "Failed to determine minimum sign bits" ); |
| 3916 | return Tmp; |
| 3917 | } |
| 3918 | } |
| 3919 | |
| 3920 | // If we are looking at the loaded value of the SDNode. |
| 3921 | if (Op.getResNo() == 0) { |
| 3922 | // Handle LOADX separately here. EXTLOAD case will fallthrough. |
| 3923 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) { |
| 3924 | unsigned ExtType = LD->getExtensionType(); |
| 3925 | switch (ExtType) { |
| 3926 | default: break; |
| 3927 | case ISD::SEXTLOAD: // e.g. i16->i32 = '17' bits known. |
| 3928 | Tmp = LD->getMemoryVT().getScalarSizeInBits(); |
| 3929 | return VTBits - Tmp + 1; |
| 3930 | case ISD::ZEXTLOAD: // e.g. i16->i32 = '16' bits known. |
| 3931 | Tmp = LD->getMemoryVT().getScalarSizeInBits(); |
| 3932 | return VTBits - Tmp; |
| 3933 | case ISD::NON_EXTLOAD: |
| 3934 | if (const Constant *Cst = TLI->getTargetConstantFromLoad(LD)) { |
| 3935 | // We only need to handle vectors - computeKnownBits should handle |
| 3936 | // scalar cases. |
| 3937 | Type *CstTy = Cst->getType(); |
| 3938 | if (CstTy->isVectorTy() && |
| 3939 | (NumElts * VTBits) == CstTy->getPrimitiveSizeInBits()) { |
| 3940 | Tmp = VTBits; |
| 3941 | for (unsigned i = 0; i != NumElts; ++i) { |
| 3942 | if (!DemandedElts[i]) |
| 3943 | continue; |
| 3944 | if (Constant *Elt = Cst->getAggregateElement(i)) { |
| 3945 | if (auto *CInt = dyn_cast<ConstantInt>(Elt)) { |
| 3946 | const APInt &Value = CInt->getValue(); |
| 3947 | Tmp = std::min(Tmp, Value.getNumSignBits()); |
| 3948 | continue; |
| 3949 | } |
| 3950 | if (auto *CFP = dyn_cast<ConstantFP>(Elt)) { |
| 3951 | APInt Value = CFP->getValueAPF().bitcastToAPInt(); |
| 3952 | Tmp = std::min(Tmp, Value.getNumSignBits()); |
| 3953 | continue; |
| 3954 | } |
| 3955 | } |
| 3956 | // Unknown type. Conservatively assume no bits match sign bit. |
| 3957 | return 1; |
| 3958 | } |
| 3959 | return Tmp; |
| 3960 | } |
| 3961 | } |
| 3962 | break; |
| 3963 | } |
| 3964 | } |
| 3965 | } |
| 3966 | |
| 3967 | // Allow the target to implement this method for its nodes. |
| 3968 | if (Opcode >= ISD::BUILTIN_OP_END || |
| 3969 | Opcode == ISD::INTRINSIC_WO_CHAIN || |
| 3970 | Opcode == ISD::INTRINSIC_W_CHAIN || |
| 3971 | Opcode == ISD::INTRINSIC_VOID) { |
| 3972 | unsigned NumBits = |
| 3973 | TLI->ComputeNumSignBitsForTargetNode(Op, DemandedElts, *this, Depth); |
| 3974 | if (NumBits > 1) |
| 3975 | FirstAnswer = std::max(FirstAnswer, NumBits); |
| 3976 | } |
| 3977 | |
| 3978 | // Finally, if we can prove that the top bits of the result are 0's or 1's, |
| 3979 | // use this information. |
| 3980 | KnownBits Known = computeKnownBits(Op, DemandedElts, Depth); |
| 3981 | |
| 3982 | APInt Mask; |
| 3983 | if (Known.isNonNegative()) { // sign bit is 0 |
| 3984 | Mask = Known.Zero; |
| 3985 | } else if (Known.isNegative()) { // sign bit is 1; |
| 3986 | Mask = Known.One; |
| 3987 | } else { |
| 3988 | // Nothing known. |
| 3989 | return FirstAnswer; |
| 3990 | } |
| 3991 | |
| 3992 | // Okay, we know that the sign bit in Mask is set. Use CLZ to determine |
| 3993 | // the number of identical bits in the top of the input value. |
| 3994 | Mask = ~Mask; |
| 3995 | Mask <<= Mask.getBitWidth()-VTBits; |
| 3996 | // Return # leading zeros. We use 'min' here in case Val was zero before |
| 3997 | // shifting. We don't want to return '64' as for an i32 "0". |
| 3998 | return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros())); |
| 3999 | } |
| 4000 | |
| 4001 | bool SelectionDAG::isBaseWithConstantOffset(SDValue Op) const { |
| 4002 | switch (Op.getOpcode()) { |
| 4003 | case ISD::ADD: |
| 4004 | case ISD::OR: |
| 4005 | case ISD::PTRADD: |
| 4006 | if (isa<ConstantSDNode>(Op.getOperand(1))) |
| 4007 | break; |
| 4008 | LLVM_FALLTHROUGH; |
| 4009 | default: |
| 4010 | return false; |
| 4011 | } |
| 4012 | |
| 4013 | if (Op.getOpcode() == ISD::OR && |
| 4014 | !MaskedValueIsZero(Op.getOperand(0), Op.getConstantOperandAPInt(1))) |
| 4015 | return false; |
| 4016 | |
| 4017 | return true; |
| 4018 | } |
| 4019 | |
| 4020 | bool SelectionDAG::isKnownNeverNaN(SDValue Op, bool SNaN, unsigned Depth) const { |
| 4021 | // If we're told that NaNs won't happen, assume they won't. |
| 4022 | if (getTarget().Options.NoNaNsFPMath || Op->getFlags().hasNoNaNs()) |
| 4023 | return true; |
| 4024 | |
| 4025 | if (Depth == 6) |
| 4026 | return false; // Limit search depth. |
| 4027 | |
| 4028 | // TODO: Handle vectors. |
| 4029 | // If the value is a constant, we can obviously see if it is a NaN or not. |
| 4030 | if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) { |
| 4031 | return !C->getValueAPF().isNaN() || |
| 4032 | (SNaN && !C->getValueAPF().isSignaling()); |
| 4033 | } |
| 4034 | |
| 4035 | unsigned Opcode = Op.getOpcode(); |
| 4036 | switch (Opcode) { |
| 4037 | case ISD::FADD: |
| 4038 | case ISD::FSUB: |
| 4039 | case ISD::FMUL: |
| 4040 | case ISD::FDIV: |
| 4041 | case ISD::FREM: |
| 4042 | case ISD::FSIN: |
| 4043 | case ISD::FCOS: { |
| 4044 | if (SNaN) |
| 4045 | return true; |
| 4046 | // TODO: Need isKnownNeverInfinity |
| 4047 | return false; |
| 4048 | } |
| 4049 | case ISD::FCANONICALIZE: |
| 4050 | case ISD::FEXP: |
| 4051 | case ISD::FEXP2: |
| 4052 | case ISD::FTRUNC: |
| 4053 | case ISD::FFLOOR: |
| 4054 | case ISD::FCEIL: |
| 4055 | case ISD::FROUND: |
| 4056 | case ISD::FRINT: |
| 4057 | case ISD::FNEARBYINT: { |
| 4058 | if (SNaN) |
| 4059 | return true; |
| 4060 | return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1); |
| 4061 | } |
| 4062 | case ISD::FABS: |
| 4063 | case ISD::FNEG: |
| 4064 | case ISD::FCOPYSIGN: { |
| 4065 | return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1); |
| 4066 | } |
| 4067 | case ISD::SELECT: |
| 4068 | return isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) && |
| 4069 | isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1); |
| 4070 | case ISD::FP_EXTEND: |
| 4071 | case ISD::FP_ROUND: { |
| 4072 | if (SNaN) |
| 4073 | return true; |
| 4074 | return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1); |
| 4075 | } |
| 4076 | case ISD::SINT_TO_FP: |
| 4077 | case ISD::UINT_TO_FP: |
| 4078 | return true; |
| 4079 | case ISD::FMA: |
| 4080 | case ISD::FMAD: { |
| 4081 | if (SNaN) |
| 4082 | return true; |
| 4083 | return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) && |
| 4084 | isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1) && |
| 4085 | isKnownNeverNaN(Op.getOperand(2), SNaN, Depth + 1); |
| 4086 | } |
| 4087 | case ISD::FSQRT: // Need is known positive |
| 4088 | case ISD::FLOG: |
| 4089 | case ISD::FLOG2: |
| 4090 | case ISD::FLOG10: |
| 4091 | case ISD::FPOWI: |
| 4092 | case ISD::FPOW: { |
| 4093 | if (SNaN) |
| 4094 | return true; |
| 4095 | // TODO: Refine on operand |
| 4096 | return false; |
| 4097 | } |
| 4098 | case ISD::FMINNUM: |
| 4099 | case ISD::FMAXNUM: { |
| 4100 | // Only one needs to be known not-nan, since it will be returned if the |
| 4101 | // other ends up being one. |
| 4102 | return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) || |
| 4103 | isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1); |
| 4104 | } |
| 4105 | case ISD::FMINNUM_IEEE: |
| 4106 | case ISD::FMAXNUM_IEEE: { |
| 4107 | if (SNaN) |
| 4108 | return true; |
| 4109 | // This can return a NaN if either operand is an sNaN, or if both operands |
| 4110 | // are NaN. |
| 4111 | return (isKnownNeverNaN(Op.getOperand(0), false, Depth + 1) && |
| 4112 | isKnownNeverSNaN(Op.getOperand(1), Depth + 1)) || |
| 4113 | (isKnownNeverNaN(Op.getOperand(1), false, Depth + 1) && |
| 4114 | isKnownNeverSNaN(Op.getOperand(0), Depth + 1)); |
| 4115 | } |
| 4116 | case ISD::FMINIMUM: |
| 4117 | case ISD::FMAXIMUM: { |
| 4118 | // TODO: Does this quiet or return the origina NaN as-is? |
| 4119 | return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) && |
| 4120 | isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1); |
| 4121 | } |
| 4122 | case ISD::EXTRACT_VECTOR_ELT: { |
| 4123 | return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1); |
| 4124 | } |
| 4125 | default: |
| 4126 | if (Opcode >= ISD::BUILTIN_OP_END || |
| 4127 | Opcode == ISD::INTRINSIC_WO_CHAIN || |
| 4128 | Opcode == ISD::INTRINSIC_W_CHAIN || |
| 4129 | Opcode == ISD::INTRINSIC_VOID) { |
| 4130 | return TLI->isKnownNeverNaNForTargetNode(Op, *this, SNaN, Depth); |
| 4131 | } |
| 4132 | |
| 4133 | return false; |
| 4134 | } |
| 4135 | } |
| 4136 | |
| 4137 | bool SelectionDAG::isKnownNeverZeroFloat(SDValue Op) const { |
| 4138 | assert(Op.getValueType().isFloatingPoint() && |
| 4139 | "Floating point type expected" ); |
| 4140 | |
| 4141 | // If the value is a constant, we can obviously see if it is a zero or not. |
| 4142 | // TODO: Add BuildVector support. |
| 4143 | if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) |
| 4144 | return !C->isZero(); |
| 4145 | return false; |
| 4146 | } |
| 4147 | |
| 4148 | bool SelectionDAG::isKnownNeverZero(SDValue Op) const { |
| 4149 | assert(!Op.getValueType().isFloatingPoint() && |
| 4150 | "Floating point types unsupported - use isKnownNeverZeroFloat" ); |
| 4151 | |
| 4152 | // If the value is a constant, we can obviously see if it is a zero or not. |
| 4153 | if (ISD::matchUnaryPredicate( |
| 4154 | Op, [](ConstantSDNode *C) { return !C->isNullValue(); })) |
| 4155 | return true; |
| 4156 | |
| 4157 | // TODO: Recognize more cases here. |
| 4158 | switch (Op.getOpcode()) { |
| 4159 | default: break; |
| 4160 | case ISD::OR: |
| 4161 | if (isKnownNeverZero(Op.getOperand(1)) || |
| 4162 | isKnownNeverZero(Op.getOperand(0))) |
| 4163 | return true; |
| 4164 | break; |
| 4165 | } |
| 4166 | |
| 4167 | return false; |
| 4168 | } |
| 4169 | |
| 4170 | bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const { |
| 4171 | // Check the obvious case. |
| 4172 | if (A == B) return true; |
| 4173 | |
| 4174 | // For for negative and positive zero. |
| 4175 | if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A)) |
| 4176 | if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B)) |
| 4177 | if (CA->isZero() && CB->isZero()) return true; |
| 4178 | |
| 4179 | // Otherwise they may not be equal. |
| 4180 | return false; |
| 4181 | } |
| 4182 | |
| 4183 | // FIXME: unify with llvm::haveNoCommonBitsSet. |
| 4184 | // FIXME: could also handle masked merge pattern (X & ~M) op (Y & M) |
| 4185 | bool SelectionDAG::haveNoCommonBitsSet(SDValue A, SDValue B) const { |
| 4186 | assert(A.getValueType() == B.getValueType() && |
| 4187 | "Values must have the same type" ); |
| 4188 | return (computeKnownBits(A).Zero | computeKnownBits(B).Zero).isAllOnesValue(); |
| 4189 | } |
| 4190 | |
| 4191 | static SDValue FoldBUILD_VECTOR(const SDLoc &DL, EVT VT, |
| 4192 | ArrayRef<SDValue> Ops, |
| 4193 | SelectionDAG &DAG) { |
| 4194 | int NumOps = Ops.size(); |
| 4195 | assert(NumOps != 0 && "Can't build an empty vector!" ); |
| 4196 | assert(VT.getVectorNumElements() == (unsigned)NumOps && |
| 4197 | "Incorrect element count in BUILD_VECTOR!" ); |
| 4198 | |
| 4199 | // BUILD_VECTOR of UNDEFs is UNDEF. |
| 4200 | if (llvm::all_of(Ops, [](SDValue Op) { return Op.isUndef(); })) |
| 4201 | return DAG.getUNDEF(VT); |
| 4202 | |
| 4203 | // BUILD_VECTOR of seq extract/insert from the same vector + type is Identity. |
| 4204 | SDValue IdentitySrc; |
| 4205 | bool IsIdentity = true; |
| 4206 | for (int i = 0; i != NumOps; ++i) { |
| 4207 | if (Ops[i].getOpcode() != ISD::EXTRACT_VECTOR_ELT || |
| 4208 | Ops[i].getOperand(0).getValueType() != VT || |
| 4209 | (IdentitySrc && Ops[i].getOperand(0) != IdentitySrc) || |
| 4210 | !isa<ConstantSDNode>(Ops[i].getOperand(1)) || |
| 4211 | cast<ConstantSDNode>(Ops[i].getOperand(1))->getAPIntValue() != i) { |
| 4212 | IsIdentity = false; |
| 4213 | break; |
| 4214 | } |
| 4215 | IdentitySrc = Ops[i].getOperand(0); |
| 4216 | } |
| 4217 | if (IsIdentity) |
| 4218 | return IdentitySrc; |
| 4219 | |
| 4220 | return SDValue(); |
| 4221 | } |
| 4222 | |
| 4223 | /// Try to simplify vector concatenation to an input value, undef, or build |
| 4224 | /// vector. |
| 4225 | static SDValue foldCONCAT_VECTORS(const SDLoc &DL, EVT VT, |
| 4226 | ArrayRef<SDValue> Ops, |
| 4227 | SelectionDAG &DAG) { |
| 4228 | assert(!Ops.empty() && "Can't concatenate an empty list of vectors!" ); |
| 4229 | assert(llvm::all_of(Ops, |
| 4230 | [Ops](SDValue Op) { |
| 4231 | return Ops[0].getValueType() == Op.getValueType(); |
| 4232 | }) && |
| 4233 | "Concatenation of vectors with inconsistent value types!" ); |
| 4234 | assert((Ops.size() * Ops[0].getValueType().getVectorNumElements()) == |
| 4235 | VT.getVectorNumElements() && |
| 4236 | "Incorrect element count in vector concatenation!" ); |
| 4237 | |
| 4238 | if (Ops.size() == 1) |
| 4239 | return Ops[0]; |
| 4240 | |
| 4241 | // Concat of UNDEFs is UNDEF. |
| 4242 | if (llvm::all_of(Ops, [](SDValue Op) { return Op.isUndef(); })) |
| 4243 | return DAG.getUNDEF(VT); |
| 4244 | |
| 4245 | // Scan the operands and look for extract operations from a single source |
| 4246 | // that correspond to insertion at the same location via this concatenation: |
| 4247 | // concat (extract X, 0*subvec_elts), (extract X, 1*subvec_elts), ... |
| 4248 | SDValue IdentitySrc; |
| 4249 | bool IsIdentity = true; |
| 4250 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) { |
| 4251 | SDValue Op = Ops[i]; |
| 4252 | unsigned IdentityIndex = i * Op.getValueType().getVectorNumElements(); |
| 4253 | if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR || |
| 4254 | Op.getOperand(0).getValueType() != VT || |
| 4255 | (IdentitySrc && Op.getOperand(0) != IdentitySrc) || |
| 4256 | !isa<ConstantSDNode>(Op.getOperand(1)) || |
| 4257 | Op.getConstantOperandVal(1) != IdentityIndex) { |
| 4258 | IsIdentity = false; |
| 4259 | break; |
| 4260 | } |
| 4261 | assert((!IdentitySrc || IdentitySrc == Op.getOperand(0)) && |
| 4262 | "Unexpected identity source vector for concat of extracts" ); |
| 4263 | IdentitySrc = Op.getOperand(0); |
| 4264 | } |
| 4265 | if (IsIdentity) { |
| 4266 | assert(IdentitySrc && "Failed to set source vector of extracts" ); |
| 4267 | return IdentitySrc; |
| 4268 | } |
| 4269 | |
| 4270 | // A CONCAT_VECTOR with all UNDEF/BUILD_VECTOR operands can be |
| 4271 | // simplified to one big BUILD_VECTOR. |
| 4272 | // FIXME: Add support for SCALAR_TO_VECTOR as well. |
| 4273 | EVT SVT = VT.getScalarType(); |
| 4274 | SmallVector<SDValue, 16> Elts; |
| 4275 | for (SDValue Op : Ops) { |
| 4276 | EVT OpVT = Op.getValueType(); |
| 4277 | if (Op.isUndef()) |
| 4278 | Elts.append(OpVT.getVectorNumElements(), DAG.getUNDEF(SVT)); |
| 4279 | else if (Op.getOpcode() == ISD::BUILD_VECTOR) |
| 4280 | Elts.append(Op->op_begin(), Op->op_end()); |
| 4281 | else |
| 4282 | return SDValue(); |
| 4283 | } |
| 4284 | |
| 4285 | // BUILD_VECTOR requires all inputs to be of the same type, find the |
| 4286 | // maximum type and extend them all. |
| 4287 | for (SDValue Op : Elts) |
| 4288 | SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT); |
| 4289 | |
| 4290 | if (SVT.bitsGT(VT.getScalarType())) |
| 4291 | for (SDValue &Op : Elts) |
| 4292 | Op = DAG.getTargetLoweringInfo().isZExtFree(Op.getValueType(), SVT) |
| 4293 | ? DAG.getZExtOrTrunc(Op, DL, SVT) |
| 4294 | : DAG.getSExtOrTrunc(Op, DL, SVT); |
| 4295 | |
| 4296 | SDValue V = DAG.getBuildVector(VT, DL, Elts); |
| 4297 | NewSDValueDbgMsg(V, "New node fold concat vectors: " , &DAG); |
| 4298 | return V; |
| 4299 | } |
| 4300 | |
| 4301 | /// Gets or creates the specified node. |
| 4302 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT) { |
| 4303 | FoldingSetNodeID ID; |
| 4304 | AddNodeIDNode(ID, Opcode, getVTList(VT), None); |
| 4305 | void *IP = nullptr; |
| 4306 | if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) |
| 4307 | return SDValue(E, 0); |
| 4308 | |
| 4309 | auto *N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), |
| 4310 | getVTList(VT)); |
| 4311 | CSEMap.InsertNode(N, IP); |
| 4312 | |
| 4313 | InsertNode(N); |
| 4314 | SDValue V = SDValue(N, 0); |
| 4315 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 4316 | return V; |
| 4317 | } |
| 4318 | |
| 4319 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, |
| 4320 | SDValue Operand, const SDNodeFlags Flags) { |
| 4321 | // Constant fold unary operations with an integer constant operand. Even |
| 4322 | // opaque constant will be folded, because the folding of unary operations |
| 4323 | // doesn't create new constants with different values. Nevertheless, the |
| 4324 | // opaque flag is preserved during folding to prevent future folding with |
| 4325 | // other constants. |
| 4326 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand)) { |
| 4327 | const APInt &Val = C->getAPIntValue(); |
| 4328 | switch (Opcode) { |
| 4329 | default: break; |
| 4330 | case ISD::SIGN_EXTEND: |
| 4331 | return getConstant(Val.sextOrTrunc(VT.getSizeInBits()), DL, VT, |
| 4332 | C->isTargetOpcode(), C->isOpaque()); |
| 4333 | case ISD::TRUNCATE: |
| 4334 | if (C->isOpaque()) |
| 4335 | break; |
| 4336 | LLVM_FALLTHROUGH; |
| 4337 | case ISD::ANY_EXTEND: |
| 4338 | case ISD::ZERO_EXTEND: |
| 4339 | return getConstant(Val.zextOrTrunc(VT.getSizeInBits()), DL, VT, |
| 4340 | C->isTargetOpcode(), C->isOpaque()); |
| 4341 | case ISD::UINT_TO_FP: |
| 4342 | case ISD::SINT_TO_FP: { |
| 4343 | APFloat apf(EVTToAPFloatSemantics(VT), |
| 4344 | APInt::getNullValue(VT.getSizeInBits())); |
| 4345 | (void)apf.convertFromAPInt(Val, |
| 4346 | Opcode==ISD::SINT_TO_FP, |
| 4347 | APFloat::rmNearestTiesToEven); |
| 4348 | return getConstantFP(apf, DL, VT); |
| 4349 | } |
| 4350 | case ISD::BITCAST: |
| 4351 | if (VT == MVT::f16 && C->getValueType(0) == MVT::i16) |
| 4352 | return getConstantFP(APFloat(APFloat::IEEEhalf(), Val), DL, VT); |
| 4353 | if (VT == MVT::f32 && C->getValueType(0) == MVT::i32) |
| 4354 | return getConstantFP(APFloat(APFloat::IEEEsingle(), Val), DL, VT); |
| 4355 | if (VT == MVT::f64 && C->getValueType(0) == MVT::i64) |
| 4356 | return getConstantFP(APFloat(APFloat::IEEEdouble(), Val), DL, VT); |
| 4357 | if (VT == MVT::f128 && C->getValueType(0) == MVT::i128) |
| 4358 | return getConstantFP(APFloat(APFloat::IEEEquad(), Val), DL, VT); |
| 4359 | break; |
| 4360 | case ISD::ABS: |
| 4361 | return getConstant(Val.abs(), DL, VT, C->isTargetOpcode(), |
| 4362 | C->isOpaque()); |
| 4363 | case ISD::BITREVERSE: |
| 4364 | return getConstant(Val.reverseBits(), DL, VT, C->isTargetOpcode(), |
| 4365 | C->isOpaque()); |
| 4366 | case ISD::BSWAP: |
| 4367 | return getConstant(Val.byteSwap(), DL, VT, C->isTargetOpcode(), |
| 4368 | C->isOpaque()); |
| 4369 | case ISD::CTPOP: |
| 4370 | return getConstant(Val.countPopulation(), DL, VT, C->isTargetOpcode(), |
| 4371 | C->isOpaque()); |
| 4372 | case ISD::CTLZ: |
| 4373 | case ISD::CTLZ_ZERO_UNDEF: |
| 4374 | return getConstant(Val.countLeadingZeros(), DL, VT, C->isTargetOpcode(), |
| 4375 | C->isOpaque()); |
| 4376 | case ISD::CTTZ: |
| 4377 | case ISD::CTTZ_ZERO_UNDEF: |
| 4378 | return getConstant(Val.countTrailingZeros(), DL, VT, C->isTargetOpcode(), |
| 4379 | C->isOpaque()); |
| 4380 | case ISD::FP16_TO_FP: { |
| 4381 | bool Ignored; |
| 4382 | APFloat FPV(APFloat::IEEEhalf(), |
| 4383 | (Val.getBitWidth() == 16) ? Val : Val.trunc(16)); |
| 4384 | |
| 4385 | // This can return overflow, underflow, or inexact; we don't care. |
| 4386 | // FIXME need to be more flexible about rounding mode. |
| 4387 | (void)FPV.convert(EVTToAPFloatSemantics(VT), |
| 4388 | APFloat::rmNearestTiesToEven, &Ignored); |
| 4389 | return getConstantFP(FPV, DL, VT); |
| 4390 | } |
| 4391 | } |
| 4392 | } |
| 4393 | |
| 4394 | // Constant fold unary operations with a floating point constant operand. |
| 4395 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand)) { |
| 4396 | APFloat V = C->getValueAPF(); // make copy |
| 4397 | switch (Opcode) { |
| 4398 | case ISD::FNEG: |
| 4399 | V.changeSign(); |
| 4400 | return getConstantFP(V, DL, VT); |
| 4401 | case ISD::FABS: |
| 4402 | V.clearSign(); |
| 4403 | return getConstantFP(V, DL, VT); |
| 4404 | case ISD::FCEIL: { |
| 4405 | APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardPositive); |
| 4406 | if (fs == APFloat::opOK || fs == APFloat::opInexact) |
| 4407 | return getConstantFP(V, DL, VT); |
| 4408 | break; |
| 4409 | } |
| 4410 | case ISD::FTRUNC: { |
| 4411 | APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardZero); |
| 4412 | if (fs == APFloat::opOK || fs == APFloat::opInexact) |
| 4413 | return getConstantFP(V, DL, VT); |
| 4414 | break; |
| 4415 | } |
| 4416 | case ISD::FFLOOR: { |
| 4417 | APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardNegative); |
| 4418 | if (fs == APFloat::opOK || fs == APFloat::opInexact) |
| 4419 | return getConstantFP(V, DL, VT); |
| 4420 | break; |
| 4421 | } |
| 4422 | case ISD::FP_EXTEND: { |
| 4423 | bool ignored; |
| 4424 | // This can return overflow, underflow, or inexact; we don't care. |
| 4425 | // FIXME need to be more flexible about rounding mode. |
| 4426 | (void)V.convert(EVTToAPFloatSemantics(VT), |
| 4427 | APFloat::rmNearestTiesToEven, &ignored); |
| 4428 | return getConstantFP(V, DL, VT); |
| 4429 | } |
| 4430 | case ISD::FP_TO_SINT: |
| 4431 | case ISD::FP_TO_UINT: { |
| 4432 | bool ignored; |
| 4433 | APSInt IntVal(VT.getSizeInBits(), Opcode == ISD::FP_TO_UINT); |
| 4434 | // FIXME need to be more flexible about rounding mode. |
| 4435 | APFloat::opStatus s = |
| 4436 | V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored); |
| 4437 | if (s == APFloat::opInvalidOp) // inexact is OK, in fact usual |
| 4438 | break; |
| 4439 | return getConstant(IntVal, DL, VT); |
| 4440 | } |
| 4441 | case ISD::BITCAST: |
| 4442 | if (VT == MVT::i16 && C->getValueType(0) == MVT::f16) |
| 4443 | return getConstant((uint16_t)V.bitcastToAPInt().getZExtValue(), DL, VT); |
| 4444 | else if (VT == MVT::i32 && C->getValueType(0) == MVT::f32) |
| 4445 | return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), DL, VT); |
| 4446 | else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64) |
| 4447 | return getConstant(V.bitcastToAPInt().getZExtValue(), DL, VT); |
| 4448 | break; |
| 4449 | case ISD::FP_TO_FP16: { |
| 4450 | bool Ignored; |
| 4451 | // This can return overflow, underflow, or inexact; we don't care. |
| 4452 | // FIXME need to be more flexible about rounding mode. |
| 4453 | (void)V.convert(APFloat::IEEEhalf(), |
| 4454 | APFloat::rmNearestTiesToEven, &Ignored); |
| 4455 | return getConstant(V.bitcastToAPInt(), DL, VT); |
| 4456 | } |
| 4457 | } |
| 4458 | } |
| 4459 | |
| 4460 | // Constant fold unary operations with a vector integer or float operand. |
| 4461 | if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Operand)) { |
| 4462 | if (BV->isConstant()) { |
| 4463 | switch (Opcode) { |
| 4464 | default: |
| 4465 | // FIXME: Entirely reasonable to perform folding of other unary |
| 4466 | // operations here as the need arises. |
| 4467 | break; |
| 4468 | case ISD::FNEG: |
| 4469 | case ISD::FABS: |
| 4470 | case ISD::FCEIL: |
| 4471 | case ISD::FTRUNC: |
| 4472 | case ISD::FFLOOR: |
| 4473 | case ISD::FP_EXTEND: |
| 4474 | case ISD::FP_TO_SINT: |
| 4475 | case ISD::FP_TO_UINT: |
| 4476 | case ISD::TRUNCATE: |
| 4477 | case ISD::ANY_EXTEND: |
| 4478 | case ISD::ZERO_EXTEND: |
| 4479 | case ISD::SIGN_EXTEND: |
| 4480 | case ISD::UINT_TO_FP: |
| 4481 | case ISD::SINT_TO_FP: |
| 4482 | case ISD::ABS: |
| 4483 | case ISD::BITREVERSE: |
| 4484 | case ISD::BSWAP: |
| 4485 | case ISD::CTLZ: |
| 4486 | case ISD::CTLZ_ZERO_UNDEF: |
| 4487 | case ISD::CTTZ: |
| 4488 | case ISD::CTTZ_ZERO_UNDEF: |
| 4489 | case ISD::CTPOP: { |
| 4490 | SDValue Ops = { Operand }; |
| 4491 | if (SDValue Fold = FoldConstantVectorArithmetic(Opcode, DL, VT, Ops)) |
| 4492 | return Fold; |
| 4493 | } |
| 4494 | } |
| 4495 | } |
| 4496 | } |
| 4497 | |
| 4498 | unsigned OpOpcode = Operand.getNode()->getOpcode(); |
| 4499 | switch (Opcode) { |
| 4500 | case ISD::TokenFactor: |
| 4501 | case ISD::MERGE_VALUES: |
| 4502 | case ISD::CONCAT_VECTORS: |
| 4503 | return Operand; // Factor, merge or concat of one node? No need. |
| 4504 | case ISD::BUILD_VECTOR: { |
| 4505 | // Attempt to simplify BUILD_VECTOR. |
| 4506 | SDValue Ops[] = {Operand}; |
| 4507 | if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this)) |
| 4508 | return V; |
| 4509 | break; |
| 4510 | } |
| 4511 | case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node" ); |
| 4512 | case ISD::FP_EXTEND: |
| 4513 | assert(VT.isFloatingPoint() && |
| 4514 | Operand.getValueType().isFloatingPoint() && "Invalid FP cast!" ); |
| 4515 | if (Operand.getValueType() == VT) return Operand; // noop conversion. |
| 4516 | assert((!VT.isVector() || |
| 4517 | VT.getVectorNumElements() == |
| 4518 | Operand.getValueType().getVectorNumElements()) && |
| 4519 | "Vector element count mismatch!" ); |
| 4520 | assert(Operand.getValueType().bitsLT(VT) && |
| 4521 | "Invalid fpext node, dst < src!" ); |
| 4522 | if (Operand.isUndef()) |
| 4523 | return getUNDEF(VT); |
| 4524 | break; |
| 4525 | case ISD::FP_TO_SINT: |
| 4526 | case ISD::FP_TO_UINT: |
| 4527 | if (Operand.isUndef()) |
| 4528 | return getUNDEF(VT); |
| 4529 | break; |
| 4530 | case ISD::SINT_TO_FP: |
| 4531 | case ISD::UINT_TO_FP: |
| 4532 | // [us]itofp(undef) = 0, because the result value is bounded. |
| 4533 | if (Operand.isUndef()) |
| 4534 | return getConstantFP(0.0, DL, VT); |
| 4535 | break; |
| 4536 | case ISD::SIGN_EXTEND: |
| 4537 | assert(VT.isInteger() && Operand.getValueType().isInteger() && |
| 4538 | "Invalid SIGN_EXTEND!" ); |
| 4539 | assert(VT.isVector() == Operand.getValueType().isVector() && |
| 4540 | "SIGN_EXTEND result type type should be vector iff the operand " |
| 4541 | "type is vector!" ); |
| 4542 | if (Operand.getValueType() == VT) return Operand; // noop extension |
| 4543 | assert((!VT.isVector() || |
| 4544 | VT.getVectorNumElements() == |
| 4545 | Operand.getValueType().getVectorNumElements()) && |
| 4546 | "Vector element count mismatch!" ); |
| 4547 | assert(Operand.getValueType().bitsLT(VT) && |
| 4548 | "Invalid sext node, dst < src!" ); |
| 4549 | if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND) |
| 4550 | return getNode(OpOpcode, DL, VT, Operand.getOperand(0)); |
| 4551 | else if (OpOpcode == ISD::UNDEF) |
| 4552 | // sext(undef) = 0, because the top bits will all be the same. |
| 4553 | return getConstant(0, DL, VT); |
| 4554 | break; |
| 4555 | case ISD::ZERO_EXTEND: |
| 4556 | assert(VT.isInteger() && Operand.getValueType().isInteger() && |
| 4557 | "Invalid ZERO_EXTEND!" ); |
| 4558 | assert(VT.isVector() == Operand.getValueType().isVector() && |
| 4559 | "ZERO_EXTEND result type type should be vector iff the operand " |
| 4560 | "type is vector!" ); |
| 4561 | if (Operand.getValueType() == VT) return Operand; // noop extension |
| 4562 | assert((!VT.isVector() || |
| 4563 | VT.getVectorNumElements() == |
| 4564 | Operand.getValueType().getVectorNumElements()) && |
| 4565 | "Vector element count mismatch!" ); |
| 4566 | assert(Operand.getValueType().bitsLT(VT) && |
| 4567 | "Invalid zext node, dst < src!" ); |
| 4568 | if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x) |
| 4569 | return getNode(ISD::ZERO_EXTEND, DL, VT, Operand.getOperand(0)); |
| 4570 | else if (OpOpcode == ISD::UNDEF) |
| 4571 | // zext(undef) = 0, because the top bits will be zero. |
| 4572 | return getConstant(0, DL, VT); |
| 4573 | break; |
| 4574 | case ISD::ANY_EXTEND: |
| 4575 | assert(VT.isInteger() && Operand.getValueType().isInteger() && |
| 4576 | "Invalid ANY_EXTEND!" ); |
| 4577 | assert(VT.isVector() == Operand.getValueType().isVector() && |
| 4578 | "ANY_EXTEND result type type should be vector iff the operand " |
| 4579 | "type is vector!" ); |
| 4580 | if (Operand.getValueType() == VT) return Operand; // noop extension |
| 4581 | assert((!VT.isVector() || |
| 4582 | VT.getVectorNumElements() == |
| 4583 | Operand.getValueType().getVectorNumElements()) && |
| 4584 | "Vector element count mismatch!" ); |
| 4585 | assert(Operand.getValueType().bitsLT(VT) && |
| 4586 | "Invalid anyext node, dst < src!" ); |
| 4587 | |
| 4588 | if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND || |
| 4589 | OpOpcode == ISD::ANY_EXTEND) |
| 4590 | // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x) |
| 4591 | return getNode(OpOpcode, DL, VT, Operand.getOperand(0)); |
| 4592 | else if (OpOpcode == ISD::UNDEF) |
| 4593 | return getUNDEF(VT); |
| 4594 | |
| 4595 | // (ext (trunc x)) -> x |
| 4596 | if (OpOpcode == ISD::TRUNCATE) { |
| 4597 | SDValue OpOp = Operand.getOperand(0); |
| 4598 | if (OpOp.getValueType() == VT) { |
| 4599 | transferDbgValues(Operand, OpOp); |
| 4600 | return OpOp; |
| 4601 | } |
| 4602 | } |
| 4603 | break; |
| 4604 | case ISD::TRUNCATE: |
| 4605 | assert(VT.isInteger() && Operand.getValueType().isInteger() && |
| 4606 | "Invalid TRUNCATE!" ); |
| 4607 | assert(VT.isVector() == Operand.getValueType().isVector() && |
| 4608 | "TRUNCATE result type type should be vector iff the operand " |
| 4609 | "type is vector!" ); |
| 4610 | if (Operand.getValueType() == VT) return Operand; // noop truncate |
| 4611 | assert((!VT.isVector() || |
| 4612 | VT.getVectorNumElements() == |
| 4613 | Operand.getValueType().getVectorNumElements()) && |
| 4614 | "Vector element count mismatch!" ); |
| 4615 | assert(Operand.getValueType().bitsGT(VT) && |
| 4616 | "Invalid truncate node, src < dst!" ); |
| 4617 | if (OpOpcode == ISD::TRUNCATE) |
| 4618 | return getNode(ISD::TRUNCATE, DL, VT, Operand.getOperand(0)); |
| 4619 | if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND || |
| 4620 | OpOpcode == ISD::ANY_EXTEND) { |
| 4621 | // If the source is smaller than the dest, we still need an extend. |
| 4622 | if (Operand.getOperand(0).getValueType().getScalarType() |
| 4623 | .bitsLT(VT.getScalarType())) |
| 4624 | return getNode(OpOpcode, DL, VT, Operand.getOperand(0)); |
| 4625 | if (Operand.getOperand(0).getValueType().bitsGT(VT)) |
| 4626 | return getNode(ISD::TRUNCATE, DL, VT, Operand.getOperand(0)); |
| 4627 | return Operand.getOperand(0); |
| 4628 | } |
| 4629 | if (OpOpcode == ISD::UNDEF) |
| 4630 | return getUNDEF(VT); |
| 4631 | break; |
| 4632 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 4633 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 4634 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 4635 | assert(VT.isVector() && "This DAG node is restricted to vector types." ); |
| 4636 | assert(Operand.getValueType().bitsLE(VT) && |
| 4637 | "The input must be the same size or smaller than the result." ); |
| 4638 | assert(VT.getVectorNumElements() < |
| 4639 | Operand.getValueType().getVectorNumElements() && |
| 4640 | "The destination vector type must have fewer lanes than the input." ); |
| 4641 | break; |
| 4642 | case ISD::ABS: |
| 4643 | assert(VT.isInteger() && VT == Operand.getValueType() && |
| 4644 | "Invalid ABS!" ); |
| 4645 | if (OpOpcode == ISD::UNDEF) |
| 4646 | return getUNDEF(VT); |
| 4647 | break; |
| 4648 | case ISD::BSWAP: |
| 4649 | assert(VT.isInteger() && VT == Operand.getValueType() && |
| 4650 | "Invalid BSWAP!" ); |
| 4651 | assert((VT.getScalarSizeInBits() % 16 == 0) && |
| 4652 | "BSWAP types must be a multiple of 16 bits!" ); |
| 4653 | if (OpOpcode == ISD::UNDEF) |
| 4654 | return getUNDEF(VT); |
| 4655 | break; |
| 4656 | case ISD::BITREVERSE: |
| 4657 | assert(VT.isInteger() && VT == Operand.getValueType() && |
| 4658 | "Invalid BITREVERSE!" ); |
| 4659 | if (OpOpcode == ISD::UNDEF) |
| 4660 | return getUNDEF(VT); |
| 4661 | break; |
| 4662 | case ISD::BITCAST: |
| 4663 | // Basic sanity checking. |
| 4664 | assert(VT.getSizeInBits() == Operand.getValueSizeInBits() && |
| 4665 | "Cannot BITCAST between types of different sizes!" ); |
| 4666 | if (VT == Operand.getValueType()) return Operand; // noop conversion. |
| 4667 | if (OpOpcode == ISD::BITCAST) // bitconv(bitconv(x)) -> bitconv(x) |
| 4668 | return getNode(ISD::BITCAST, DL, VT, Operand.getOperand(0)); |
| 4669 | if (OpOpcode == ISD::UNDEF) |
| 4670 | return getUNDEF(VT); |
| 4671 | break; |
| 4672 | case ISD::SCALAR_TO_VECTOR: |
| 4673 | assert(VT.isVector() && !Operand.getValueType().isVector() && |
| 4674 | (VT.getVectorElementType() == Operand.getValueType() || |
| 4675 | (VT.getVectorElementType().isInteger() && |
| 4676 | Operand.getValueType().isInteger() && |
| 4677 | VT.getVectorElementType().bitsLE(Operand.getValueType()))) && |
| 4678 | "Illegal SCALAR_TO_VECTOR node!" ); |
| 4679 | if (OpOpcode == ISD::UNDEF) |
| 4680 | return getUNDEF(VT); |
| 4681 | // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined. |
| 4682 | if (OpOpcode == ISD::EXTRACT_VECTOR_ELT && |
| 4683 | isa<ConstantSDNode>(Operand.getOperand(1)) && |
| 4684 | Operand.getConstantOperandVal(1) == 0 && |
| 4685 | Operand.getOperand(0).getValueType() == VT) |
| 4686 | return Operand.getOperand(0); |
| 4687 | break; |
| 4688 | case ISD::FNEG: |
| 4689 | // Negation of an unknown bag of bits is still completely undefined. |
| 4690 | if (OpOpcode == ISD::UNDEF) |
| 4691 | return getUNDEF(VT); |
| 4692 | |
| 4693 | // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0 |
| 4694 | if ((getTarget().Options.UnsafeFPMath || Flags.hasNoSignedZeros()) && |
| 4695 | OpOpcode == ISD::FSUB) |
| 4696 | return getNode(ISD::FSUB, DL, VT, Operand.getOperand(1), |
| 4697 | Operand.getOperand(0), Flags); |
| 4698 | if (OpOpcode == ISD::FNEG) // --X -> X |
| 4699 | return Operand.getOperand(0); |
| 4700 | break; |
| 4701 | case ISD::FABS: |
| 4702 | if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X) |
| 4703 | return getNode(ISD::FABS, DL, VT, Operand.getOperand(0)); |
| 4704 | break; |
| 4705 | } |
| 4706 | |
| 4707 | SDNode *N; |
| 4708 | SDVTList VTs = getVTList(VT); |
| 4709 | SDValue Ops[] = {Operand}; |
| 4710 | if (VT != MVT::Glue) { // Don't CSE flag producing nodes |
| 4711 | FoldingSetNodeID ID; |
| 4712 | AddNodeIDNode(ID, Opcode, VTs, Ops); |
| 4713 | void *IP = nullptr; |
| 4714 | if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) { |
| 4715 | E->intersectFlagsWith(Flags); |
| 4716 | return SDValue(E, 0); |
| 4717 | } |
| 4718 | |
| 4719 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 4720 | N->setFlags(Flags); |
| 4721 | createOperands(N, Ops); |
| 4722 | CSEMap.InsertNode(N, IP); |
| 4723 | } else { |
| 4724 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 4725 | createOperands(N, Ops); |
| 4726 | } |
| 4727 | |
| 4728 | InsertNode(N); |
| 4729 | SDValue V = SDValue(N, 0); |
| 4730 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 4731 | return V; |
| 4732 | } |
| 4733 | |
| 4734 | static std::pair<APInt, bool> FoldValue(unsigned Opcode, const APInt &C1, |
| 4735 | const APInt &C2) { |
| 4736 | switch (Opcode) { |
| 4737 | case ISD::ADD: return std::make_pair(C1 + C2, true); |
| 4738 | case ISD::SUB: return std::make_pair(C1 - C2, true); |
| 4739 | case ISD::MUL: return std::make_pair(C1 * C2, true); |
| 4740 | case ISD::AND: return std::make_pair(C1 & C2, true); |
| 4741 | case ISD::OR: return std::make_pair(C1 | C2, true); |
| 4742 | case ISD::XOR: return std::make_pair(C1 ^ C2, true); |
| 4743 | case ISD::SHL: return std::make_pair(C1 << C2, true); |
| 4744 | case ISD::SRL: return std::make_pair(C1.lshr(C2), true); |
| 4745 | case ISD::SRA: return std::make_pair(C1.ashr(C2), true); |
| 4746 | case ISD::ROTL: return std::make_pair(C1.rotl(C2), true); |
| 4747 | case ISD::ROTR: return std::make_pair(C1.rotr(C2), true); |
| 4748 | case ISD::SMIN: return std::make_pair(C1.sle(C2) ? C1 : C2, true); |
| 4749 | case ISD::SMAX: return std::make_pair(C1.sge(C2) ? C1 : C2, true); |
| 4750 | case ISD::UMIN: return std::make_pair(C1.ule(C2) ? C1 : C2, true); |
| 4751 | case ISD::UMAX: return std::make_pair(C1.uge(C2) ? C1 : C2, true); |
| 4752 | case ISD::SADDSAT: return std::make_pair(C1.sadd_sat(C2), true); |
| 4753 | case ISD::UADDSAT: return std::make_pair(C1.uadd_sat(C2), true); |
| 4754 | case ISD::SSUBSAT: return std::make_pair(C1.ssub_sat(C2), true); |
| 4755 | case ISD::USUBSAT: return std::make_pair(C1.usub_sat(C2), true); |
| 4756 | case ISD::UDIV: |
| 4757 | if (!C2.getBoolValue()) |
| 4758 | break; |
| 4759 | return std::make_pair(C1.udiv(C2), true); |
| 4760 | case ISD::UREM: |
| 4761 | if (!C2.getBoolValue()) |
| 4762 | break; |
| 4763 | return std::make_pair(C1.urem(C2), true); |
| 4764 | case ISD::SDIV: |
| 4765 | if (!C2.getBoolValue()) |
| 4766 | break; |
| 4767 | return std::make_pair(C1.sdiv(C2), true); |
| 4768 | case ISD::SREM: |
| 4769 | if (!C2.getBoolValue()) |
| 4770 | break; |
| 4771 | return std::make_pair(C1.srem(C2), true); |
| 4772 | } |
| 4773 | return std::make_pair(APInt(1, 0), false); |
| 4774 | } |
| 4775 | |
| 4776 | SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, const SDLoc &DL, |
| 4777 | EVT VT, const ConstantSDNode *C1, |
| 4778 | const ConstantSDNode *C2) { |
| 4779 | if (C1->isOpaque() || C2->isOpaque()) |
| 4780 | return SDValue(); |
| 4781 | |
| 4782 | std::pair<APInt, bool> Folded = FoldValue(Opcode, C1->getAPIntValue(), |
| 4783 | C2->getAPIntValue()); |
| 4784 | if (!Folded.second) |
| 4785 | return SDValue(); |
| 4786 | return getConstant(Folded.first, DL, VT); |
| 4787 | } |
| 4788 | |
| 4789 | SDValue SelectionDAG::FoldSymbolOffset(unsigned Opcode, EVT VT, |
| 4790 | const GlobalAddressSDNode *GA, |
| 4791 | const SDNode *N2) { |
| 4792 | if (GA->getOpcode() != ISD::GlobalAddress) |
| 4793 | return SDValue(); |
| 4794 | if (!TLI->isOffsetFoldingLegal(GA)) |
| 4795 | return SDValue(); |
| 4796 | auto *C2 = dyn_cast<ConstantSDNode>(N2); |
| 4797 | if (!C2) |
| 4798 | return SDValue(); |
| 4799 | int64_t Offset = C2->getSExtValue(); |
| 4800 | switch (Opcode) { |
| 4801 | case ISD::ADD: break; |
| 4802 | case ISD::SUB: Offset = -uint64_t(Offset); break; |
| 4803 | default: return SDValue(); |
| 4804 | } |
| 4805 | return getGlobalAddress(GA->getGlobal(), SDLoc(C2), VT, |
| 4806 | GA->getOffset() + uint64_t(Offset)); |
| 4807 | } |
| 4808 | |
| 4809 | bool SelectionDAG::isUndef(unsigned Opcode, ArrayRef<SDValue> Ops) { |
| 4810 | switch (Opcode) { |
| 4811 | case ISD::SDIV: |
| 4812 | case ISD::UDIV: |
| 4813 | case ISD::SREM: |
| 4814 | case ISD::UREM: { |
| 4815 | // If a divisor is zero/undef or any element of a divisor vector is |
| 4816 | // zero/undef, the whole op is undef. |
| 4817 | assert(Ops.size() == 2 && "Div/rem should have 2 operands" ); |
| 4818 | SDValue Divisor = Ops[1]; |
| 4819 | if (Divisor.isUndef() || isNullConstant(Divisor)) |
| 4820 | return true; |
| 4821 | |
| 4822 | return ISD::isBuildVectorOfConstantSDNodes(Divisor.getNode()) && |
| 4823 | llvm::any_of(Divisor->op_values(), |
| 4824 | [](SDValue V) { return V.isUndef() || |
| 4825 | isNullConstant(V); }); |
| 4826 | // TODO: Handle signed overflow. |
| 4827 | } |
| 4828 | // TODO: Handle oversized shifts. |
| 4829 | default: |
| 4830 | return false; |
| 4831 | } |
| 4832 | } |
| 4833 | |
| 4834 | SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, const SDLoc &DL, |
| 4835 | EVT VT, SDNode *N1, SDNode *N2) { |
| 4836 | // If the opcode is a target-specific ISD node, there's nothing we can |
| 4837 | // do here and the operand rules may not line up with the below, so |
| 4838 | // bail early. |
| 4839 | if (Opcode >= ISD::BUILTIN_OP_END) |
| 4840 | return SDValue(); |
| 4841 | |
| 4842 | if (isUndef(Opcode, {SDValue(N1, 0), SDValue(N2, 0)})) |
| 4843 | return getUNDEF(VT); |
| 4844 | |
| 4845 | // Handle the case of two scalars. |
| 4846 | if (auto *C1 = dyn_cast<ConstantSDNode>(N1)) { |
| 4847 | if (auto *C2 = dyn_cast<ConstantSDNode>(N2)) { |
| 4848 | SDValue Folded = FoldConstantArithmetic(Opcode, DL, VT, C1, C2); |
| 4849 | assert((!Folded || !VT.isVector()) && |
| 4850 | "Can't fold vectors ops with scalar operands" ); |
| 4851 | return Folded; |
| 4852 | } |
| 4853 | } |
| 4854 | |
| 4855 | // fold (add Sym, c) -> Sym+c |
| 4856 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N1)) |
| 4857 | return FoldSymbolOffset(Opcode, VT, GA, N2); |
| 4858 | if (TLI->isCommutativeBinOp(Opcode)) |
| 4859 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N2)) |
| 4860 | return FoldSymbolOffset(Opcode, VT, GA, N1); |
| 4861 | |
| 4862 | // For vectors, extract each constant element and fold them individually. |
| 4863 | // Either input may be an undef value. |
| 4864 | auto *BV1 = dyn_cast<BuildVectorSDNode>(N1); |
| 4865 | if (!BV1 && !N1->isUndef()) |
| 4866 | return SDValue(); |
| 4867 | auto *BV2 = dyn_cast<BuildVectorSDNode>(N2); |
| 4868 | if (!BV2 && !N2->isUndef()) |
| 4869 | return SDValue(); |
| 4870 | // If both operands are undef, that's handled the same way as scalars. |
| 4871 | if (!BV1 && !BV2) |
| 4872 | return SDValue(); |
| 4873 | |
| 4874 | assert((!BV1 || !BV2 || BV1->getNumOperands() == BV2->getNumOperands()) && |
| 4875 | "Vector binop with different number of elements in operands?" ); |
| 4876 | |
| 4877 | EVT SVT = VT.getScalarType(); |
| 4878 | EVT LegalSVT = SVT; |
| 4879 | if (NewNodesMustHaveLegalTypes && LegalSVT.isInteger()) { |
| 4880 | LegalSVT = TLI->getTypeToTransformTo(*getContext(), LegalSVT); |
| 4881 | if (LegalSVT.bitsLT(SVT)) |
| 4882 | return SDValue(); |
| 4883 | } |
| 4884 | SmallVector<SDValue, 4> Outputs; |
| 4885 | unsigned NumOps = BV1 ? BV1->getNumOperands() : BV2->getNumOperands(); |
| 4886 | for (unsigned I = 0; I != NumOps; ++I) { |
| 4887 | SDValue V1 = BV1 ? BV1->getOperand(I) : getUNDEF(SVT); |
| 4888 | SDValue V2 = BV2 ? BV2->getOperand(I) : getUNDEF(SVT); |
| 4889 | if (SVT.isInteger()) { |
| 4890 | if (V1->getValueType(0).bitsGT(SVT)) |
| 4891 | V1 = getNode(ISD::TRUNCATE, DL, SVT, V1); |
| 4892 | if (V2->getValueType(0).bitsGT(SVT)) |
| 4893 | V2 = getNode(ISD::TRUNCATE, DL, SVT, V2); |
| 4894 | } |
| 4895 | |
| 4896 | if (V1->getValueType(0) != SVT || V2->getValueType(0) != SVT) |
| 4897 | return SDValue(); |
| 4898 | |
| 4899 | // Fold one vector element. |
| 4900 | SDValue ScalarResult = getNode(Opcode, DL, SVT, V1, V2); |
| 4901 | if (LegalSVT != SVT) |
| 4902 | ScalarResult = getNode(ISD::SIGN_EXTEND, DL, LegalSVT, ScalarResult); |
| 4903 | |
| 4904 | // Scalar folding only succeeded if the result is a constant or UNDEF. |
| 4905 | if (!ScalarResult.isUndef() && ScalarResult.getOpcode() != ISD::Constant && |
| 4906 | ScalarResult.getOpcode() != ISD::ConstantFP) |
| 4907 | return SDValue(); |
| 4908 | Outputs.push_back(ScalarResult); |
| 4909 | } |
| 4910 | |
| 4911 | assert(VT.getVectorNumElements() == Outputs.size() && |
| 4912 | "Vector size mismatch!" ); |
| 4913 | |
| 4914 | // We may have a vector type but a scalar result. Create a splat. |
| 4915 | Outputs.resize(VT.getVectorNumElements(), Outputs.back()); |
| 4916 | |
| 4917 | // Build a big vector out of the scalar elements we generated. |
| 4918 | return getBuildVector(VT, SDLoc(), Outputs); |
| 4919 | } |
| 4920 | |
| 4921 | // TODO: Merge with FoldConstantArithmetic |
| 4922 | SDValue SelectionDAG::FoldConstantVectorArithmetic(unsigned Opcode, |
| 4923 | const SDLoc &DL, EVT VT, |
| 4924 | ArrayRef<SDValue> Ops, |
| 4925 | const SDNodeFlags Flags) { |
| 4926 | // If the opcode is a target-specific ISD node, there's nothing we can |
| 4927 | // do here and the operand rules may not line up with the below, so |
| 4928 | // bail early. |
| 4929 | if (Opcode >= ISD::BUILTIN_OP_END) |
| 4930 | return SDValue(); |
| 4931 | |
| 4932 | if (isUndef(Opcode, Ops)) |
| 4933 | return getUNDEF(VT); |
| 4934 | |
| 4935 | // We can only fold vectors - maybe merge with FoldConstantArithmetic someday? |
| 4936 | if (!VT.isVector()) |
| 4937 | return SDValue(); |
| 4938 | |
| 4939 | unsigned NumElts = VT.getVectorNumElements(); |
| 4940 | |
| 4941 | auto IsScalarOrSameVectorSize = [&](const SDValue &Op) { |
| 4942 | return !Op.getValueType().isVector() || |
| 4943 | Op.getValueType().getVectorNumElements() == NumElts; |
| 4944 | }; |
| 4945 | |
| 4946 | auto IsConstantBuildVectorOrUndef = [&](const SDValue &Op) { |
| 4947 | BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op); |
| 4948 | return (Op.isUndef()) || (Op.getOpcode() == ISD::CONDCODE) || |
| 4949 | (BV && BV->isConstant()); |
| 4950 | }; |
| 4951 | |
| 4952 | // All operands must be vector types with the same number of elements as |
| 4953 | // the result type and must be either UNDEF or a build vector of constant |
| 4954 | // or UNDEF scalars. |
| 4955 | if (!llvm::all_of(Ops, IsConstantBuildVectorOrUndef) || |
| 4956 | !llvm::all_of(Ops, IsScalarOrSameVectorSize)) |
| 4957 | return SDValue(); |
| 4958 | |
| 4959 | // If we are comparing vectors, then the result needs to be a i1 boolean |
| 4960 | // that is then sign-extended back to the legal result type. |
| 4961 | EVT SVT = (Opcode == ISD::SETCC ? MVT::i1 : VT.getScalarType()); |
| 4962 | |
| 4963 | // Find legal integer scalar type for constant promotion and |
| 4964 | // ensure that its scalar size is at least as large as source. |
| 4965 | EVT LegalSVT = VT.getScalarType(); |
| 4966 | if (NewNodesMustHaveLegalTypes && LegalSVT.isInteger()) { |
| 4967 | LegalSVT = TLI->getTypeToTransformTo(*getContext(), LegalSVT); |
| 4968 | if (LegalSVT.bitsLT(VT.getScalarType())) |
| 4969 | return SDValue(); |
| 4970 | } |
| 4971 | |
| 4972 | // Constant fold each scalar lane separately. |
| 4973 | SmallVector<SDValue, 4> ScalarResults; |
| 4974 | for (unsigned i = 0; i != NumElts; i++) { |
| 4975 | SmallVector<SDValue, 4> ScalarOps; |
| 4976 | for (SDValue Op : Ops) { |
| 4977 | EVT InSVT = Op.getValueType().getScalarType(); |
| 4978 | BuildVectorSDNode *InBV = dyn_cast<BuildVectorSDNode>(Op); |
| 4979 | if (!InBV) { |
| 4980 | // We've checked that this is UNDEF or a constant of some kind. |
| 4981 | if (Op.isUndef()) |
| 4982 | ScalarOps.push_back(getUNDEF(InSVT)); |
| 4983 | else |
| 4984 | ScalarOps.push_back(Op); |
| 4985 | continue; |
| 4986 | } |
| 4987 | |
| 4988 | SDValue ScalarOp = InBV->getOperand(i); |
| 4989 | EVT ScalarVT = ScalarOp.getValueType(); |
| 4990 | |
| 4991 | // Build vector (integer) scalar operands may need implicit |
| 4992 | // truncation - do this before constant folding. |
| 4993 | if (ScalarVT.isInteger() && ScalarVT.bitsGT(InSVT)) |
| 4994 | ScalarOp = getNode(ISD::TRUNCATE, DL, InSVT, ScalarOp); |
| 4995 | |
| 4996 | ScalarOps.push_back(ScalarOp); |
| 4997 | } |
| 4998 | |
| 4999 | // Constant fold the scalar operands. |
| 5000 | SDValue ScalarResult = getNode(Opcode, DL, SVT, ScalarOps, Flags); |
| 5001 | |
| 5002 | // Legalize the (integer) scalar constant if necessary. |
| 5003 | if (LegalSVT != SVT) |
| 5004 | ScalarResult = getNode(ISD::SIGN_EXTEND, DL, LegalSVT, ScalarResult); |
| 5005 | |
| 5006 | // Scalar folding only succeeded if the result is a constant or UNDEF. |
| 5007 | if (!ScalarResult.isUndef() && ScalarResult.getOpcode() != ISD::Constant && |
| 5008 | ScalarResult.getOpcode() != ISD::ConstantFP) |
| 5009 | return SDValue(); |
| 5010 | ScalarResults.push_back(ScalarResult); |
| 5011 | } |
| 5012 | |
| 5013 | SDValue V = getBuildVector(VT, DL, ScalarResults); |
| 5014 | NewSDValueDbgMsg(V, "New node fold constant vector: " , this); |
| 5015 | return V; |
| 5016 | } |
| 5017 | |
| 5018 | SDValue SelectionDAG::foldConstantFPMath(unsigned Opcode, const SDLoc &DL, |
| 5019 | EVT VT, SDValue N1, SDValue N2) { |
| 5020 | // TODO: We don't do any constant folding for strict FP opcodes here, but we |
| 5021 | // should. That will require dealing with a potentially non-default |
| 5022 | // rounding mode, checking the "opStatus" return value from the APFloat |
| 5023 | // math calculations, and possibly other variations. |
| 5024 | auto *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode()); |
| 5025 | auto *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode()); |
| 5026 | if (N1CFP && N2CFP) { |
| 5027 | APFloat C1 = N1CFP->getValueAPF(), C2 = N2CFP->getValueAPF(); |
| 5028 | switch (Opcode) { |
| 5029 | case ISD::FADD: |
| 5030 | C1.add(C2, APFloat::rmNearestTiesToEven); |
| 5031 | return getConstantFP(C1, DL, VT); |
| 5032 | case ISD::FSUB: |
| 5033 | C1.subtract(C2, APFloat::rmNearestTiesToEven); |
| 5034 | return getConstantFP(C1, DL, VT); |
| 5035 | case ISD::FMUL: |
| 5036 | C1.multiply(C2, APFloat::rmNearestTiesToEven); |
| 5037 | return getConstantFP(C1, DL, VT); |
| 5038 | case ISD::FDIV: |
| 5039 | C1.divide(C2, APFloat::rmNearestTiesToEven); |
| 5040 | return getConstantFP(C1, DL, VT); |
| 5041 | case ISD::FREM: |
| 5042 | C1.mod(C2); |
| 5043 | return getConstantFP(C1, DL, VT); |
| 5044 | case ISD::FCOPYSIGN: |
| 5045 | C1.copySign(C2); |
| 5046 | return getConstantFP(C1, DL, VT); |
| 5047 | default: break; |
| 5048 | } |
| 5049 | } |
| 5050 | if (N1CFP && Opcode == ISD::FP_ROUND) { |
| 5051 | APFloat C1 = N1CFP->getValueAPF(); // make copy |
| 5052 | bool Unused; |
| 5053 | // This can return overflow, underflow, or inexact; we don't care. |
| 5054 | // FIXME need to be more flexible about rounding mode. |
| 5055 | (void) C1.convert(EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven, |
| 5056 | &Unused); |
| 5057 | return getConstantFP(C1, DL, VT); |
| 5058 | } |
| 5059 | |
| 5060 | switch (Opcode) { |
| 5061 | case ISD::FADD: |
| 5062 | case ISD::FSUB: |
| 5063 | case ISD::FMUL: |
| 5064 | case ISD::FDIV: |
| 5065 | case ISD::FREM: |
| 5066 | // If both operands are undef, the result is undef. If 1 operand is undef, |
| 5067 | // the result is NaN. This should match the behavior of the IR optimizer. |
| 5068 | if (N1.isUndef() && N2.isUndef()) |
| 5069 | return getUNDEF(VT); |
| 5070 | if (N1.isUndef() || N2.isUndef()) |
| 5071 | return getConstantFP(APFloat::getNaN(EVTToAPFloatSemantics(VT)), DL, VT); |
| 5072 | } |
| 5073 | return SDValue(); |
| 5074 | } |
| 5075 | |
| 5076 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, |
| 5077 | SDValue N1, SDValue N2, const SDNodeFlags Flags) { |
| 5078 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); |
| 5079 | ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); |
| 5080 | ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); |
| 5081 | ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2); |
| 5082 | |
| 5083 | // Canonicalize constant to RHS if commutative. |
| 5084 | if (TLI->isCommutativeBinOp(Opcode)) { |
| 5085 | if (N1C && !N2C) { |
| 5086 | std::swap(N1C, N2C); |
| 5087 | std::swap(N1, N2); |
| 5088 | } else if (N1CFP && !N2CFP) { |
| 5089 | std::swap(N1CFP, N2CFP); |
| 5090 | std::swap(N1, N2); |
| 5091 | } |
| 5092 | } |
| 5093 | |
| 5094 | switch (Opcode) { |
| 5095 | default: break; |
| 5096 | case ISD::TokenFactor: |
| 5097 | assert(VT == MVT::Other && N1.getValueType() == MVT::Other && |
| 5098 | N2.getValueType() == MVT::Other && "Invalid token factor!" ); |
| 5099 | // Fold trivial token factors. |
| 5100 | if (N1.getOpcode() == ISD::EntryToken) return N2; |
| 5101 | if (N2.getOpcode() == ISD::EntryToken) return N1; |
| 5102 | if (N1 == N2) return N1; |
| 5103 | break; |
| 5104 | case ISD::BUILD_VECTOR: { |
| 5105 | // Attempt to simplify BUILD_VECTOR. |
| 5106 | SDValue Ops[] = {N1, N2}; |
| 5107 | if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this)) |
| 5108 | return V; |
| 5109 | break; |
| 5110 | } |
| 5111 | case ISD::CONCAT_VECTORS: { |
| 5112 | SDValue Ops[] = {N1, N2}; |
| 5113 | if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, *this)) |
| 5114 | return V; |
| 5115 | break; |
| 5116 | } |
| 5117 | case ISD::AND: |
| 5118 | assert(VT.isInteger() && "This operator does not apply to FP types!" ); |
| 5119 | assert(N1.getValueType() == N2.getValueType() && |
| 5120 | N1.getValueType() == VT && "Binary operator types must match!" ); |
| 5121 | // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's |
| 5122 | // worth handling here. |
| 5123 | if (N2C && N2C->isNullValue()) |
| 5124 | return N2; |
| 5125 | if (N2C && N2C->isAllOnesValue()) // X & -1 -> X |
| 5126 | return N1; |
| 5127 | break; |
| 5128 | case ISD::OR: |
| 5129 | case ISD::XOR: |
| 5130 | case ISD::ADD: |
| 5131 | case ISD::SUB: |
| 5132 | assert(!VT.isFatPointer() && |
| 5133 | "This operator does not apply to capability types!" ); |
| 5134 | assert(VT.isInteger() && "This operator does not apply to FP types!" ); |
| 5135 | assert(N1.getValueType() == N2.getValueType() && |
| 5136 | N1.getValueType() == VT && "Binary operator types must match!" ); |
| 5137 | // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so |
| 5138 | // it's worth handling here. |
| 5139 | if (N2C && N2C->isNullValue()) |
| 5140 | return N1; |
| 5141 | break; |
| 5142 | case ISD::UDIV: |
| 5143 | case ISD::UREM: |
| 5144 | case ISD::MULHU: |
| 5145 | case ISD::MULHS: |
| 5146 | case ISD::MUL: |
| 5147 | case ISD::SDIV: |
| 5148 | case ISD::SREM: |
| 5149 | case ISD::SMIN: |
| 5150 | case ISD::SMAX: |
| 5151 | case ISD::UMIN: |
| 5152 | case ISD::UMAX: |
| 5153 | case ISD::SADDSAT: |
| 5154 | case ISD::SSUBSAT: |
| 5155 | case ISD::UADDSAT: |
| 5156 | case ISD::USUBSAT: |
| 5157 | assert(VT.isInteger() && "This operator does not apply to FP types!" ); |
| 5158 | assert(N1.getValueType() == N2.getValueType() && |
| 5159 | N1.getValueType() == VT && "Binary operator types must match!" ); |
| 5160 | break; |
| 5161 | case ISD::FADD: |
| 5162 | case ISD::FSUB: |
| 5163 | case ISD::FMUL: |
| 5164 | case ISD::FDIV: |
| 5165 | case ISD::FREM: |
| 5166 | assert(VT.isFloatingPoint() && "This operator only applies to FP types!" ); |
| 5167 | assert(N1.getValueType() == N2.getValueType() && |
| 5168 | N1.getValueType() == VT && "Binary operator types must match!" ); |
| 5169 | if (SDValue V = simplifyFPBinop(Opcode, N1, N2)) |
| 5170 | return V; |
| 5171 | break; |
| 5172 | case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match. |
| 5173 | assert(N1.getValueType() == VT && |
| 5174 | N1.getValueType().isFloatingPoint() && |
| 5175 | N2.getValueType().isFloatingPoint() && |
| 5176 | "Invalid FCOPYSIGN!" ); |
| 5177 | break; |
| 5178 | case ISD::SHL: |
| 5179 | case ISD::SRA: |
| 5180 | case ISD::SRL: |
| 5181 | if (SDValue V = simplifyShift(N1, N2)) |
| 5182 | return V; |
| 5183 | LLVM_FALLTHROUGH; |
| 5184 | case ISD::ROTL: |
| 5185 | case ISD::ROTR: |
| 5186 | assert(VT == N1.getValueType() && |
| 5187 | "Shift operators return type must be the same as their first arg" ); |
| 5188 | assert(VT.isInteger() && N2.getValueType().isInteger() && |
| 5189 | "Shifts only work on integers" ); |
| 5190 | assert((!VT.isVector() || VT == N2.getValueType()) && |
| 5191 | "Vector shift amounts must be in the same as their first arg" ); |
| 5192 | // Verify that the shift amount VT is big enough to hold valid shift |
| 5193 | // amounts. This catches things like trying to shift an i1024 value by an |
| 5194 | // i8, which is easy to fall into in generic code that uses |
| 5195 | // TLI.getShiftAmount(). |
| 5196 | assert(N2.getValueSizeInBits() >= Log2_32_Ceil(N1.getValueSizeInBits()) && |
| 5197 | "Invalid use of small shift amount with oversized value!" ); |
| 5198 | |
| 5199 | // Always fold shifts of i1 values so the code generator doesn't need to |
| 5200 | // handle them. Since we know the size of the shift has to be less than the |
| 5201 | // size of the value, the shift/rotate count is guaranteed to be zero. |
| 5202 | if (VT == MVT::i1) |
| 5203 | return N1; |
| 5204 | if (N2C && N2C->isNullValue()) |
| 5205 | return N1; |
| 5206 | break; |
| 5207 | case ISD::FP_ROUND_INREG: { |
| 5208 | EVT EVT = cast<VTSDNode>(N2)->getVT(); |
| 5209 | assert(VT == N1.getValueType() && "Not an inreg round!" ); |
| 5210 | assert(VT.isFloatingPoint() && EVT.isFloatingPoint() && |
| 5211 | "Cannot FP_ROUND_INREG integer types" ); |
| 5212 | assert(EVT.isVector() == VT.isVector() && |
| 5213 | "FP_ROUND_INREG type should be vector iff the operand " |
| 5214 | "type is vector!" ); |
| 5215 | assert((!EVT.isVector() || |
| 5216 | EVT.getVectorNumElements() == VT.getVectorNumElements()) && |
| 5217 | "Vector element counts must match in FP_ROUND_INREG" ); |
| 5218 | assert(EVT.bitsLE(VT) && "Not rounding down!" ); |
| 5219 | (void)EVT; |
| 5220 | if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding. |
| 5221 | break; |
| 5222 | } |
| 5223 | case ISD::FP_ROUND: |
| 5224 | assert(VT.isFloatingPoint() && |
| 5225 | N1.getValueType().isFloatingPoint() && |
| 5226 | VT.bitsLE(N1.getValueType()) && |
| 5227 | N2C && (N2C->getZExtValue() == 0 || N2C->getZExtValue() == 1) && |
| 5228 | "Invalid FP_ROUND!" ); |
| 5229 | if (N1.getValueType() == VT) return N1; // noop conversion. |
| 5230 | break; |
| 5231 | case ISD::AssertSext: |
| 5232 | case ISD::AssertZext: { |
| 5233 | EVT EVT = cast<VTSDNode>(N2)->getVT(); |
| 5234 | assert(VT == N1.getValueType() && "Not an inreg extend!" ); |
| 5235 | assert(VT.isInteger() && EVT.isInteger() && |
| 5236 | "Cannot *_EXTEND_INREG FP types" ); |
| 5237 | assert(!EVT.isVector() && |
| 5238 | "AssertSExt/AssertZExt type should be the vector element type " |
| 5239 | "rather than the vector type!" ); |
| 5240 | assert(EVT.bitsLE(VT.getScalarType()) && "Not extending!" ); |
| 5241 | if (VT.getScalarType() == EVT) return N1; // noop assertion. |
| 5242 | break; |
| 5243 | } |
| 5244 | case ISD::SIGN_EXTEND_INREG: { |
| 5245 | EVT EVT = cast<VTSDNode>(N2)->getVT(); |
| 5246 | assert(VT == N1.getValueType() && "Not an inreg extend!" ); |
| 5247 | assert(VT.isInteger() && EVT.isInteger() && |
| 5248 | "Cannot *_EXTEND_INREG FP types" ); |
| 5249 | assert(EVT.isVector() == VT.isVector() && |
| 5250 | "SIGN_EXTEND_INREG type should be vector iff the operand " |
| 5251 | "type is vector!" ); |
| 5252 | assert((!EVT.isVector() || |
| 5253 | EVT.getVectorNumElements() == VT.getVectorNumElements()) && |
| 5254 | "Vector element counts must match in SIGN_EXTEND_INREG" ); |
| 5255 | assert(EVT.bitsLE(VT) && "Not extending!" ); |
| 5256 | if (EVT == VT) return N1; // Not actually extending |
| 5257 | |
| 5258 | auto SignExtendInReg = [&](APInt Val, llvm::EVT ConstantVT) { |
| 5259 | unsigned FromBits = EVT.getScalarSizeInBits(); |
| 5260 | Val <<= Val.getBitWidth() - FromBits; |
| 5261 | Val.ashrInPlace(Val.getBitWidth() - FromBits); |
| 5262 | return getConstant(Val, DL, ConstantVT); |
| 5263 | }; |
| 5264 | |
| 5265 | if (N1C) { |
| 5266 | const APInt &Val = N1C->getAPIntValue(); |
| 5267 | return SignExtendInReg(Val, VT); |
| 5268 | } |
| 5269 | if (ISD::isBuildVectorOfConstantSDNodes(N1.getNode())) { |
| 5270 | SmallVector<SDValue, 8> Ops; |
| 5271 | llvm::EVT OpVT = N1.getOperand(0).getValueType(); |
| 5272 | for (int i = 0, e = VT.getVectorNumElements(); i != e; ++i) { |
| 5273 | SDValue Op = N1.getOperand(i); |
| 5274 | if (Op.isUndef()) { |
| 5275 | Ops.push_back(getUNDEF(OpVT)); |
| 5276 | continue; |
| 5277 | } |
| 5278 | ConstantSDNode *C = cast<ConstantSDNode>(Op); |
| 5279 | APInt Val = C->getAPIntValue(); |
| 5280 | Ops.push_back(SignExtendInReg(Val, OpVT)); |
| 5281 | } |
| 5282 | return getBuildVector(VT, DL, Ops); |
| 5283 | } |
| 5284 | break; |
| 5285 | } |
| 5286 | case ISD::EXTRACT_VECTOR_ELT: |
| 5287 | assert(VT.getSizeInBits() >= N1.getValueType().getScalarSizeInBits() && |
| 5288 | "The result of EXTRACT_VECTOR_ELT must be at least as wide as the \ |
| 5289 | element type of the vector." ); |
| 5290 | |
| 5291 | // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF. |
| 5292 | if (N1.isUndef()) |
| 5293 | return getUNDEF(VT); |
| 5294 | |
| 5295 | // EXTRACT_VECTOR_ELT of out-of-bounds element is an UNDEF |
| 5296 | if (N2C && N2C->getAPIntValue().uge(N1.getValueType().getVectorNumElements())) |
| 5297 | return getUNDEF(VT); |
| 5298 | |
| 5299 | // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is |
| 5300 | // expanding copies of large vectors from registers. |
| 5301 | if (N2C && |
| 5302 | N1.getOpcode() == ISD::CONCAT_VECTORS && |
| 5303 | N1.getNumOperands() > 0) { |
| 5304 | unsigned Factor = |
| 5305 | N1.getOperand(0).getValueType().getVectorNumElements(); |
| 5306 | return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, |
| 5307 | N1.getOperand(N2C->getZExtValue() / Factor), |
| 5308 | getConstant(N2C->getZExtValue() % Factor, DL, |
| 5309 | N2.getValueType())); |
| 5310 | } |
| 5311 | |
| 5312 | // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is |
| 5313 | // expanding large vector constants. |
| 5314 | if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) { |
| 5315 | SDValue Elt = N1.getOperand(N2C->getZExtValue()); |
| 5316 | |
| 5317 | if (VT != Elt.getValueType()) |
| 5318 | // If the vector element type is not legal, the BUILD_VECTOR operands |
| 5319 | // are promoted and implicitly truncated, and the result implicitly |
| 5320 | // extended. Make that explicit here. |
| 5321 | Elt = getAnyExtOrTrunc(Elt, DL, VT); |
| 5322 | |
| 5323 | return Elt; |
| 5324 | } |
| 5325 | |
| 5326 | // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector |
| 5327 | // operations are lowered to scalars. |
| 5328 | if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) { |
| 5329 | // If the indices are the same, return the inserted element else |
| 5330 | // if the indices are known different, extract the element from |
| 5331 | // the original vector. |
| 5332 | SDValue N1Op2 = N1.getOperand(2); |
| 5333 | ConstantSDNode *N1Op2C = dyn_cast<ConstantSDNode>(N1Op2); |
| 5334 | |
| 5335 | if (N1Op2C && N2C) { |
| 5336 | if (N1Op2C->getZExtValue() == N2C->getZExtValue()) { |
| 5337 | if (VT == N1.getOperand(1).getValueType()) |
| 5338 | return N1.getOperand(1); |
| 5339 | else |
| 5340 | return getSExtOrTrunc(N1.getOperand(1), DL, VT); |
| 5341 | } |
| 5342 | |
| 5343 | return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2); |
| 5344 | } |
| 5345 | } |
| 5346 | |
| 5347 | // EXTRACT_VECTOR_ELT of v1iX EXTRACT_SUBVECTOR could be formed |
| 5348 | // when vector types are scalarized and v1iX is legal. |
| 5349 | // vextract (v1iX extract_subvector(vNiX, Idx)) -> vextract(vNiX,Idx) |
| 5350 | if (N1.getOpcode() == ISD::EXTRACT_SUBVECTOR && |
| 5351 | N1.getValueType().getVectorNumElements() == 1) { |
| 5352 | return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), |
| 5353 | N1.getOperand(1)); |
| 5354 | } |
| 5355 | break; |
| 5356 | case ISD::EXTRACT_ELEMENT: |
| 5357 | assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!" ); |
| 5358 | assert(!N1.getValueType().isVector() && !VT.isVector() && |
| 5359 | (N1.getValueType().isInteger() == VT.isInteger()) && |
| 5360 | N1.getValueType() != VT && |
| 5361 | "Wrong types for EXTRACT_ELEMENT!" ); |
| 5362 | |
| 5363 | // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding |
| 5364 | // 64-bit integers into 32-bit parts. Instead of building the extract of |
| 5365 | // the BUILD_PAIR, only to have legalize rip it apart, just do it now. |
| 5366 | if (N1.getOpcode() == ISD::BUILD_PAIR) |
| 5367 | return N1.getOperand(N2C->getZExtValue()); |
| 5368 | |
| 5369 | // EXTRACT_ELEMENT of a constant int is also very common. |
| 5370 | if (N1C) { |
| 5371 | unsigned ElementSize = VT.getSizeInBits(); |
| 5372 | unsigned Shift = ElementSize * N2C->getZExtValue(); |
| 5373 | APInt ShiftedVal = N1C->getAPIntValue().lshr(Shift); |
| 5374 | return getConstant(ShiftedVal.trunc(ElementSize), DL, VT); |
| 5375 | } |
| 5376 | break; |
| 5377 | case ISD::EXTRACT_SUBVECTOR: |
| 5378 | if (VT.isSimple() && N1.getValueType().isSimple()) { |
| 5379 | assert(VT.isVector() && N1.getValueType().isVector() && |
| 5380 | "Extract subvector VTs must be a vectors!" ); |
| 5381 | assert(VT.getVectorElementType() == |
| 5382 | N1.getValueType().getVectorElementType() && |
| 5383 | "Extract subvector VTs must have the same element type!" ); |
| 5384 | assert(VT.getSimpleVT() <= N1.getSimpleValueType() && |
| 5385 | "Extract subvector must be from larger vector to smaller vector!" ); |
| 5386 | |
| 5387 | if (N2C) { |
| 5388 | assert((VT.getVectorNumElements() + N2C->getZExtValue() |
| 5389 | <= N1.getValueType().getVectorNumElements()) |
| 5390 | && "Extract subvector overflow!" ); |
| 5391 | } |
| 5392 | |
| 5393 | // Trivial extraction. |
| 5394 | if (VT.getSimpleVT() == N1.getSimpleValueType()) |
| 5395 | return N1; |
| 5396 | |
| 5397 | // EXTRACT_SUBVECTOR of an UNDEF is an UNDEF. |
| 5398 | if (N1.isUndef()) |
| 5399 | return getUNDEF(VT); |
| 5400 | |
| 5401 | // EXTRACT_SUBVECTOR of CONCAT_VECTOR can be simplified if the pieces of |
| 5402 | // the concat have the same type as the extract. |
| 5403 | if (N2C && N1.getOpcode() == ISD::CONCAT_VECTORS && |
| 5404 | N1.getNumOperands() > 0 && |
| 5405 | VT == N1.getOperand(0).getValueType()) { |
| 5406 | unsigned Factor = VT.getVectorNumElements(); |
| 5407 | return N1.getOperand(N2C->getZExtValue() / Factor); |
| 5408 | } |
| 5409 | |
| 5410 | // EXTRACT_SUBVECTOR of INSERT_SUBVECTOR is often created |
| 5411 | // during shuffle legalization. |
| 5412 | if (N1.getOpcode() == ISD::INSERT_SUBVECTOR && N2 == N1.getOperand(2) && |
| 5413 | VT == N1.getOperand(1).getValueType()) |
| 5414 | return N1.getOperand(1); |
| 5415 | } |
| 5416 | break; |
| 5417 | } |
| 5418 | |
| 5419 | // Perform trivial constant folding. |
| 5420 | if (SDValue SV = |
| 5421 | FoldConstantArithmetic(Opcode, DL, VT, N1.getNode(), N2.getNode())) |
| 5422 | return SV; |
| 5423 | |
| 5424 | if (SDValue V = foldConstantFPMath(Opcode, DL, VT, N1, N2)) |
| 5425 | return V; |
| 5426 | |
| 5427 | // Canonicalize an UNDEF to the RHS, even over a constant. |
| 5428 | if (N1.isUndef()) { |
| 5429 | if (TLI->isCommutativeBinOp(Opcode)) { |
| 5430 | std::swap(N1, N2); |
| 5431 | } else { |
| 5432 | switch (Opcode) { |
| 5433 | case ISD::FP_ROUND_INREG: |
| 5434 | case ISD::SIGN_EXTEND_INREG: |
| 5435 | case ISD::SUB: |
| 5436 | return getUNDEF(VT); // fold op(undef, arg2) -> undef |
| 5437 | case ISD::UDIV: |
| 5438 | case ISD::SDIV: |
| 5439 | case ISD::UREM: |
| 5440 | case ISD::SREM: |
| 5441 | case ISD::SSUBSAT: |
| 5442 | case ISD::USUBSAT: |
| 5443 | return getConstant(0, DL, VT); // fold op(undef, arg2) -> 0 |
| 5444 | } |
| 5445 | } |
| 5446 | } |
| 5447 | |
| 5448 | // Fold a bunch of operators when the RHS is undef. |
| 5449 | if (N2.isUndef()) { |
| 5450 | switch (Opcode) { |
| 5451 | case ISD::XOR: |
| 5452 | if (N1.isUndef()) |
| 5453 | // Handle undef ^ undef -> 0 special case. This is a common |
| 5454 | // idiom (misuse). |
| 5455 | return getConstant(0, DL, VT); |
| 5456 | LLVM_FALLTHROUGH; |
| 5457 | case ISD::ADD: |
| 5458 | case ISD::SUB: |
| 5459 | case ISD::UDIV: |
| 5460 | case ISD::SDIV: |
| 5461 | case ISD::UREM: |
| 5462 | case ISD::SREM: |
| 5463 | return getUNDEF(VT); // fold op(arg1, undef) -> undef |
| 5464 | case ISD::MUL: |
| 5465 | case ISD::AND: |
| 5466 | case ISD::SSUBSAT: |
| 5467 | case ISD::USUBSAT: |
| 5468 | return getConstant(0, DL, VT); // fold op(arg1, undef) -> 0 |
| 5469 | case ISD::OR: |
| 5470 | case ISD::SADDSAT: |
| 5471 | case ISD::UADDSAT: |
| 5472 | return getAllOnesConstant(DL, VT); |
| 5473 | } |
| 5474 | } |
| 5475 | |
| 5476 | // Memoize this node if possible. |
| 5477 | SDNode *N; |
| 5478 | SDVTList VTs = getVTList(VT); |
| 5479 | SDValue Ops[] = {N1, N2}; |
| 5480 | if (VT != MVT::Glue) { |
| 5481 | FoldingSetNodeID ID; |
| 5482 | AddNodeIDNode(ID, Opcode, VTs, Ops); |
| 5483 | void *IP = nullptr; |
| 5484 | if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) { |
| 5485 | E->intersectFlagsWith(Flags); |
| 5486 | return SDValue(E, 0); |
| 5487 | } |
| 5488 | |
| 5489 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 5490 | N->setFlags(Flags); |
| 5491 | createOperands(N, Ops); |
| 5492 | CSEMap.InsertNode(N, IP); |
| 5493 | } else { |
| 5494 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 5495 | createOperands(N, Ops); |
| 5496 | } |
| 5497 | |
| 5498 | InsertNode(N); |
| 5499 | SDValue V = SDValue(N, 0); |
| 5500 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 5501 | return V; |
| 5502 | } |
| 5503 | |
| 5504 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, |
| 5505 | SDValue N1, SDValue N2, SDValue N3, |
| 5506 | const SDNodeFlags Flags) { |
| 5507 | // Perform various simplifications. |
| 5508 | switch (Opcode) { |
| 5509 | case ISD::FMA: { |
| 5510 | assert(VT.isFloatingPoint() && "This operator only applies to FP types!" ); |
| 5511 | assert(N1.getValueType() == VT && N2.getValueType() == VT && |
| 5512 | N3.getValueType() == VT && "FMA types must match!" ); |
| 5513 | ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); |
| 5514 | ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2); |
| 5515 | ConstantFPSDNode *N3CFP = dyn_cast<ConstantFPSDNode>(N3); |
| 5516 | if (N1CFP && N2CFP && N3CFP) { |
| 5517 | APFloat V1 = N1CFP->getValueAPF(); |
| 5518 | const APFloat &V2 = N2CFP->getValueAPF(); |
| 5519 | const APFloat &V3 = N3CFP->getValueAPF(); |
| 5520 | V1.fusedMultiplyAdd(V2, V3, APFloat::rmNearestTiesToEven); |
| 5521 | return getConstantFP(V1, DL, VT); |
| 5522 | } |
| 5523 | break; |
| 5524 | } |
| 5525 | case ISD::BUILD_VECTOR: { |
| 5526 | // Attempt to simplify BUILD_VECTOR. |
| 5527 | SDValue Ops[] = {N1, N2, N3}; |
| 5528 | if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this)) |
| 5529 | return V; |
| 5530 | break; |
| 5531 | } |
| 5532 | case ISD::CONCAT_VECTORS: { |
| 5533 | SDValue Ops[] = {N1, N2, N3}; |
| 5534 | if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, *this)) |
| 5535 | return V; |
| 5536 | break; |
| 5537 | } |
| 5538 | case ISD::SETCC: { |
| 5539 | assert(VT.isInteger() && "SETCC result type must be an integer!" ); |
| 5540 | assert(N1.getValueType() == N2.getValueType() && |
| 5541 | "SETCC operands must have the same type!" ); |
| 5542 | assert(VT.isVector() == N1.getValueType().isVector() && |
| 5543 | "SETCC type should be vector iff the operand type is vector!" ); |
| 5544 | assert((!VT.isVector() || |
| 5545 | VT.getVectorNumElements() == N1.getValueType().getVectorNumElements()) && |
| 5546 | "SETCC vector element counts must match!" ); |
| 5547 | // Use FoldSetCC to simplify SETCC's. |
| 5548 | if (SDValue V = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL)) |
| 5549 | return V; |
| 5550 | // Vector constant folding. |
| 5551 | SDValue Ops[] = {N1, N2, N3}; |
| 5552 | if (SDValue V = FoldConstantVectorArithmetic(Opcode, DL, VT, Ops)) { |
| 5553 | NewSDValueDbgMsg(V, "New node vector constant folding: " , this); |
| 5554 | return V; |
| 5555 | } |
| 5556 | break; |
| 5557 | } |
| 5558 | case ISD::SELECT: |
| 5559 | case ISD::VSELECT: |
| 5560 | if (SDValue V = simplifySelect(N1, N2, N3)) |
| 5561 | return V; |
| 5562 | break; |
| 5563 | case ISD::VECTOR_SHUFFLE: |
| 5564 | llvm_unreachable("should use getVectorShuffle constructor!" ); |
| 5565 | case ISD::INSERT_VECTOR_ELT: { |
| 5566 | ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3); |
| 5567 | // INSERT_VECTOR_ELT into out-of-bounds element is an UNDEF |
| 5568 | if (N3C && N3C->getZExtValue() >= N1.getValueType().getVectorNumElements()) |
| 5569 | return getUNDEF(VT); |
| 5570 | break; |
| 5571 | } |
| 5572 | case ISD::INSERT_SUBVECTOR: { |
| 5573 | // Inserting undef into undef is still undef. |
| 5574 | if (N1.isUndef() && N2.isUndef()) |
| 5575 | return getUNDEF(VT); |
| 5576 | SDValue Index = N3; |
| 5577 | if (VT.isSimple() && N1.getValueType().isSimple() |
| 5578 | && N2.getValueType().isSimple()) { |
| 5579 | assert(VT.isVector() && N1.getValueType().isVector() && |
| 5580 | N2.getValueType().isVector() && |
| 5581 | "Insert subvector VTs must be a vectors" ); |
| 5582 | assert(VT == N1.getValueType() && |
| 5583 | "Dest and insert subvector source types must match!" ); |
| 5584 | assert(N2.getSimpleValueType() <= N1.getSimpleValueType() && |
| 5585 | "Insert subvector must be from smaller vector to larger vector!" ); |
| 5586 | if (isa<ConstantSDNode>(Index)) { |
| 5587 | assert((N2.getValueType().getVectorNumElements() + |
| 5588 | cast<ConstantSDNode>(Index)->getZExtValue() |
| 5589 | <= VT.getVectorNumElements()) |
| 5590 | && "Insert subvector overflow!" ); |
| 5591 | } |
| 5592 | |
| 5593 | // Trivial insertion. |
| 5594 | if (VT.getSimpleVT() == N2.getSimpleValueType()) |
| 5595 | return N2; |
| 5596 | } |
| 5597 | break; |
| 5598 | } |
| 5599 | case ISD::BITCAST: |
| 5600 | // Fold bit_convert nodes from a type to themselves. |
| 5601 | if (N1.getValueType() == VT) |
| 5602 | return N1; |
| 5603 | break; |
| 5604 | } |
| 5605 | |
| 5606 | // Memoize node if it doesn't produce a flag. |
| 5607 | SDNode *N; |
| 5608 | SDVTList VTs = getVTList(VT); |
| 5609 | SDValue Ops[] = {N1, N2, N3}; |
| 5610 | if (VT != MVT::Glue) { |
| 5611 | FoldingSetNodeID ID; |
| 5612 | AddNodeIDNode(ID, Opcode, VTs, Ops); |
| 5613 | void *IP = nullptr; |
| 5614 | if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) { |
| 5615 | E->intersectFlagsWith(Flags); |
| 5616 | return SDValue(E, 0); |
| 5617 | } |
| 5618 | |
| 5619 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 5620 | N->setFlags(Flags); |
| 5621 | createOperands(N, Ops); |
| 5622 | CSEMap.InsertNode(N, IP); |
| 5623 | } else { |
| 5624 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 5625 | createOperands(N, Ops); |
| 5626 | } |
| 5627 | |
| 5628 | InsertNode(N); |
| 5629 | SDValue V = SDValue(N, 0); |
| 5630 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 5631 | return V; |
| 5632 | } |
| 5633 | |
| 5634 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, |
| 5635 | SDValue N1, SDValue N2, SDValue N3, SDValue N4) { |
| 5636 | SDValue Ops[] = { N1, N2, N3, N4 }; |
| 5637 | return getNode(Opcode, DL, VT, Ops); |
| 5638 | } |
| 5639 | |
| 5640 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, |
| 5641 | SDValue N1, SDValue N2, SDValue N3, SDValue N4, |
| 5642 | SDValue N5) { |
| 5643 | SDValue Ops[] = { N1, N2, N3, N4, N5 }; |
| 5644 | return getNode(Opcode, DL, VT, Ops); |
| 5645 | } |
| 5646 | |
| 5647 | /// getStackArgumentTokenFactor - Compute a TokenFactor to force all |
| 5648 | /// the incoming stack arguments to be loaded from the stack. |
| 5649 | SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) { |
| 5650 | SmallVector<SDValue, 8> ArgChains; |
| 5651 | |
| 5652 | // Include the original chain at the beginning of the list. When this is |
| 5653 | // used by target LowerCall hooks, this helps legalize find the |
| 5654 | // CALLSEQ_BEGIN node. |
| 5655 | ArgChains.push_back(Chain); |
| 5656 | |
| 5657 | // Add a chain value for each stack argument. |
| 5658 | for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(), |
| 5659 | UE = getEntryNode().getNode()->use_end(); U != UE; ++U) |
| 5660 | if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U)) |
| 5661 | if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) |
| 5662 | if (FI->getIndex() < 0) |
| 5663 | ArgChains.push_back(SDValue(L, 1)); |
| 5664 | |
| 5665 | // Build a tokenfactor for all the chains. |
| 5666 | return getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains); |
| 5667 | } |
| 5668 | |
| 5669 | /// getMemsetValue - Vectorized representation of the memset value |
| 5670 | /// operand. |
| 5671 | static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG, |
| 5672 | const SDLoc &dl) { |
| 5673 | assert(!Value.isUndef()); |
| 5674 | |
| 5675 | unsigned NumBits = VT.getScalarSizeInBits(); |
| 5676 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) { |
| 5677 | assert(C->getAPIntValue().getBitWidth() == 8); |
| 5678 | APInt Val = APInt::getSplat(NumBits, C->getAPIntValue()); |
| 5679 | if (VT.isInteger() || VT.isFatPointer()) { |
| 5680 | bool IsOpaque = VT.getSizeInBits() > 64 || |
| 5681 | !DAG.getTargetLoweringInfo().isLegalStoreImmediate(C->getSExtValue()); |
| 5682 | return DAG.getConstant(Val, dl, VT, false, IsOpaque); |
| 5683 | } |
| 5684 | return DAG.getConstantFP(APFloat(DAG.EVTToAPFloatSemantics(VT), Val), dl, |
| 5685 | VT); |
| 5686 | } |
| 5687 | |
| 5688 | assert(Value.getValueType() == MVT::i8 && "memset with non-byte fill value?" ); |
| 5689 | EVT IntVT = VT.getScalarType(); |
| 5690 | if (!IntVT.isInteger()) |
| 5691 | IntVT = EVT::getIntegerVT(*DAG.getContext(), IntVT.getSizeInBits()); |
| 5692 | |
| 5693 | Value = DAG.getNode(ISD::ZERO_EXTEND, dl, IntVT, Value); |
| 5694 | if (NumBits > 8) { |
| 5695 | // Use a multiplication with 0x010101... to extend the input to the |
| 5696 | // required length. |
| 5697 | APInt Magic = APInt::getSplat(NumBits, APInt(8, 0x01)); |
| 5698 | Value = DAG.getNode(ISD::MUL, dl, IntVT, Value, |
| 5699 | DAG.getConstant(Magic, dl, IntVT)); |
| 5700 | } |
| 5701 | |
| 5702 | if (VT != Value.getValueType() && !VT.isInteger()) |
| 5703 | Value = DAG.getBitcast(VT.getScalarType(), Value); |
| 5704 | if (VT != Value.getValueType()) |
| 5705 | Value = DAG.getSplatBuildVector(VT, dl, Value); |
| 5706 | |
| 5707 | return Value; |
| 5708 | } |
| 5709 | |
| 5710 | /// getMemsetStringVal - Similar to getMemsetValue. Except this is only |
| 5711 | /// used when a memcpy is turned into a memset when the source is a constant |
| 5712 | /// string ptr. |
| 5713 | static SDValue getMemsetStringVal(EVT VT, const SDLoc &dl, SelectionDAG &DAG, |
| 5714 | const TargetLowering &TLI, |
| 5715 | const ConstantDataArraySlice &Slice) { |
| 5716 | // Handle vector with all elements zero. |
| 5717 | if (Slice.Array == nullptr) { |
| 5718 | if (VT.isFatPointer()) |
| 5719 | return DAG.getNullCapability(dl, VT); |
| 5720 | if (VT.isInteger()) |
| 5721 | return DAG.getConstant(0, dl, VT); |
| 5722 | else if (VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128) |
| 5723 | return DAG.getConstantFP(0.0, dl, VT); |
| 5724 | else if (VT.isVector()) { |
| 5725 | unsigned NumElts = VT.getVectorNumElements(); |
| 5726 | MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64; |
| 5727 | return DAG.getNode(ISD::BITCAST, dl, VT, |
| 5728 | DAG.getConstant(0, dl, |
| 5729 | EVT::getVectorVT(*DAG.getContext(), |
| 5730 | EltVT, NumElts))); |
| 5731 | } else |
| 5732 | llvm_unreachable("Expected type!" ); |
| 5733 | } |
| 5734 | |
| 5735 | assert(!VT.isVector() && "Can't handle vector type here!" ); |
| 5736 | unsigned NumVTBits = VT.getSizeInBits(); |
| 5737 | unsigned NumVTBytes = NumVTBits / 8; |
| 5738 | unsigned NumBytes = std::min(NumVTBytes, unsigned(Slice.Length)); |
| 5739 | |
| 5740 | APInt Val(NumVTBits, 0); |
| 5741 | if (DAG.getDataLayout().isLittleEndian()) { |
| 5742 | for (unsigned i = 0; i != NumBytes; ++i) |
| 5743 | Val |= (uint64_t)(unsigned char)Slice[i] << i*8; |
| 5744 | } else { |
| 5745 | for (unsigned i = 0; i != NumBytes; ++i) |
| 5746 | Val |= (uint64_t)(unsigned char)Slice[i] << (NumVTBytes-i-1)*8; |
| 5747 | } |
| 5748 | |
| 5749 | // If the "cost" of materializing the integer immediate is less than the cost |
| 5750 | // of a load, then it is cost effective to turn the load into the immediate. |
| 5751 | Type *Ty = VT.getTypeForEVT(*DAG.getContext()); |
| 5752 | if (TLI.shouldConvertConstantLoadToIntImm(Val, Ty)) |
| 5753 | return DAG.getConstant(Val, dl, VT); |
| 5754 | return SDValue(nullptr, 0); |
| 5755 | } |
| 5756 | |
| 5757 | /// getMemBasePlusOffset - Returns base and offset node for the |
| 5758 | /// |
| 5759 | SDValue SelectionDAG::getMemBasePlusOffset(SDValue Base, unsigned Offset, |
| 5760 | const SDLoc &DL, |
| 5761 | const SDNodeFlags Flags) { |
| 5762 | return getPointerAdd(DL, Base, Offset, Flags); |
| 5763 | } |
| 5764 | |
| 5765 | /// Returns true if memcpy source is constant data. |
| 5766 | static bool isMemSrcFromConstant(SDValue Src, ConstantDataArraySlice &Slice) { |
| 5767 | uint64_t SrcDelta = 0; |
| 5768 | GlobalAddressSDNode *G = nullptr; |
| 5769 | if (Src.getOpcode() == ISD::GlobalAddress) |
| 5770 | G = cast<GlobalAddressSDNode>(Src); |
| 5771 | else if (Src.getOpcode() == ISD::ADD && |
| 5772 | Src.getOperand(0).getOpcode() == ISD::GlobalAddress && |
| 5773 | Src.getOperand(1).getOpcode() == ISD::Constant) { |
| 5774 | G = cast<GlobalAddressSDNode>(Src.getOperand(0)); |
| 5775 | SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue(); |
| 5776 | } |
| 5777 | if (!G) |
| 5778 | return false; |
| 5779 | |
| 5780 | return getConstantDataArrayInfo(G->getGlobal(), Slice, 8, |
| 5781 | SrcDelta + G->getOffset()); |
| 5782 | } |
| 5783 | |
| 5784 | static bool shouldLowerMemFuncForSize(const MachineFunction &MF) { |
| 5785 | // On Darwin, -Os means optimize for size without hurting performance, so |
| 5786 | // only really optimize for size when -Oz (MinSize) is used. |
| 5787 | if (MF.getTarget().getTargetTriple().isOSDarwin()) |
| 5788 | return MF.getFunction().hasMinSize(); |
| 5789 | return MF.getFunction().hasOptSize(); |
| 5790 | } |
| 5791 | |
| 5792 | static void chainLoadsAndStoresForMemcpy(SelectionDAG &DAG, const SDLoc &dl, |
| 5793 | SmallVector<SDValue, 32> &OutChains, unsigned From, |
| 5794 | unsigned To, SmallVector<SDValue, 16> &OutLoadChains, |
| 5795 | SmallVector<SDValue, 16> &OutStoreChains) { |
| 5796 | assert(OutLoadChains.size() && "Missing loads in memcpy inlining" ); |
| 5797 | assert(OutStoreChains.size() && "Missing stores in memcpy inlining" ); |
| 5798 | SmallVector<SDValue, 16> GluedLoadChains; |
| 5799 | for (unsigned i = From; i < To; ++i) { |
| 5800 | OutChains.push_back(OutLoadChains[i]); |
| 5801 | GluedLoadChains.push_back(OutLoadChains[i]); |
| 5802 | } |
| 5803 | |
| 5804 | // Chain for all loads. |
| 5805 | SDValue LoadToken = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, |
| 5806 | GluedLoadChains); |
| 5807 | |
| 5808 | for (unsigned i = From; i < To; ++i) { |
| 5809 | StoreSDNode *ST = dyn_cast<StoreSDNode>(OutStoreChains[i]); |
| 5810 | SDValue NewStore = DAG.getTruncStore(LoadToken, dl, ST->getValue(), |
| 5811 | ST->getBasePtr(), ST->getMemoryVT(), |
| 5812 | ST->getMemOperand()); |
| 5813 | OutChains.push_back(NewStore); |
| 5814 | } |
| 5815 | } |
| 5816 | |
| 5817 | static void |
| 5818 | diagnoseInefficientCheriMemOp(SelectionDAG &DAG, const DiagnosticLocation &Loc, |
| 5819 | const Twine &MemOp, CodeGenOpt::Level OptLevel, |
| 5820 | StringRef Type, unsigned Align, uint64_t Size, |
| 5821 | uint64_t CapSize) { |
| 5822 | assert(Align < CapSize); |
| 5823 | assert(Size >= CapSize); |
| 5824 | if (OptLevel == CodeGenOpt::None) |
| 5825 | return; // Don't bother warning about inefficient code at -O0 |
| 5826 | // Skip the memcpy/memmove diag if we have already diagnosed something else |
| 5827 | if (Type == "!!<CHERI-NODIAG>!!" ) |
| 5828 | return; |
| 5829 | |
| 5830 | DiagnosticInfoCheriInefficient Warning( |
| 5831 | DAG.getMachineFunction().getFunction(), Loc, |
| 5832 | MemOp + " operation with capability argument " + Type + |
| 5833 | " and underaligned destination (aligned to " + Twine(Align) + |
| 5834 | " bytes) may be inefficient or result in CHERI tags bits being " |
| 5835 | "stripped" ); |
| 5836 | DAG.getContext()->diagnose(Warning); |
| 5837 | } |
| 5838 | |
| 5839 | static SDValue getMemcpyLoadsAndStores( |
| 5840 | SelectionDAG &DAG, const SDLoc &dl, SDValue Chain, SDValue Dst, SDValue Src, |
| 5841 | uint64_t Size, unsigned Align, bool isVol, bool AlwaysInline, |
| 5842 | bool MustPreserveCheriCapabilities, MachinePointerInfo DstPtrInfo, |
| 5843 | MachinePointerInfo SrcPtrInfo, StringRef CopyTy, |
| 5844 | CodeGenOpt::Level OptLevel) { |
| 5845 | // Turn a memcpy of undef to nop. |
| 5846 | if (Src.isUndef()) |
| 5847 | return Chain; |
| 5848 | |
| 5849 | // Expand memcpy to a series of load and store ops if the size operand falls |
| 5850 | // below a certain threshold. |
| 5851 | // TODO: In the AlwaysInline case, if the size is big then generate a loop |
| 5852 | // rather than maybe a humongous number of loads and stores. |
| 5853 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
| 5854 | const DataLayout &DL = DAG.getDataLayout(); |
| 5855 | LLVMContext &C = *DAG.getContext(); |
| 5856 | std::vector<EVT> MemOps; |
| 5857 | bool DstAlignCanChange = false; |
| 5858 | MachineFunction &MF = DAG.getMachineFunction(); |
| 5859 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 5860 | bool OptSize = shouldLowerMemFuncForSize(MF); |
| 5861 | FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst); |
| 5862 | if (FI && !MFI.isFixedObjectIndex(FI->getIndex())) |
| 5863 | DstAlignCanChange = true; |
| 5864 | unsigned SrcAlign = DAG.InferPtrAlignment(Src); |
| 5865 | if (Align > SrcAlign) |
| 5866 | SrcAlign = Align; |
| 5867 | ConstantDataArraySlice Slice; |
| 5868 | bool CopyFromConstant = isMemSrcFromConstant(Src, Slice); |
| 5869 | bool isZeroConstant = CopyFromConstant && Slice.Array == nullptr; |
| 5870 | unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemcpy(OptSize); |
| 5871 | const bool FoundLowering = TLI.findOptimalMemOpLowering(MemOps, Limit, Size, |
| 5872 | (DstAlignCanChange ? 0 : Align), |
| 5873 | (isZeroConstant ? 0 : SrcAlign), |
| 5874 | false, false, CopyFromConstant, true, MustPreserveCheriCapabilities, |
| 5875 | DstPtrInfo.getAddrSpace(), |
| 5876 | SrcPtrInfo.getAddrSpace(), |
| 5877 | MF.getFunction().getAttributes()); |
| 5878 | |
| 5879 | // Don't warn about inefficient memcpy if we reached the inline memcpy limit |
| 5880 | // Also don't warn about copies of less than CapSize |
| 5881 | // TODO: the frontend probably shouldn't emit must-preserve-tags for such |
| 5882 | // small memcpys |
| 5883 | auto CapTy = TLI.cheriCapabilityType(); |
| 5884 | const uint64_t CapSize = CapTy.isValid() ? CapTy.getStoreSize() : 0; |
| 5885 | bool ReachedLimit = (CapSize * Limit) < Size; |
| 5886 | if (MustPreserveCheriCapabilities && !ReachedLimit && Size >= CapSize && |
| 5887 | (!FoundLowering || !MemOps[0].isFatPointer())) { |
| 5888 | LLVM_DEBUG( |
| 5889 | dbgs() << " memcpy must preserve tags but value is not statically " |
| 5890 | "known to be sufficiently aligned -> using memcpy() call\n" ); |
| 5891 | if (AlwaysInline) { |
| 5892 | report_fatal_error("MustPreserveCheriCapabilities and AlwaysInline set " |
| 5893 | "but operation cannot be lowered to loads+stores!" ); |
| 5894 | } |
| 5895 | diagnoseInefficientCheriMemOp(DAG, dl.getDebugLoc(), "memcpy" , OptLevel, |
| 5896 | CopyTy.empty() ? "<unknown type>" : CopyTy, |
| 5897 | std::max(1u, std::min(Align, SrcAlign)), |
| 5898 | Size, CapSize); |
| 5899 | return SDValue(); |
| 5900 | } |
| 5901 | if (!FoundLowering) |
| 5902 | return SDValue(); |
| 5903 | |
| 5904 | if (DstAlignCanChange) { |
| 5905 | Type *Ty = MemOps[0].getTypeForEVT(C); |
| 5906 | LLVM_DEBUG(dbgs() << " DstAlignCanChange -> using type " ; Ty->dump()); |
| 5907 | unsigned NewAlign = (unsigned)DL.getABITypeAlignment(Ty); |
| 5908 | LLVM_DEBUG(dbgs() << "\t->NewAlign = " << NewAlign << ", stack alignment=" |
| 5909 | << DL.getStackAlignment() << "\n" ); |
| 5910 | if (MemOps[0].isFatPointer()) { |
| 5911 | assert(!DL.exceedsNaturalStackAlignment(NewAlign) && |
| 5912 | "Stack not capability-aligned?" ); |
| 5913 | } |
| 5914 | |
| 5915 | // Don't promote to an alignment that would require dynamic stack |
| 5916 | // realignment. |
| 5917 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
| 5918 | if (!TRI->needsStackRealignment(MF)) |
| 5919 | while (NewAlign > Align && |
| 5920 | DL.exceedsNaturalStackAlignment(NewAlign)) |
| 5921 | NewAlign /= 2; |
| 5922 | if (MemOps[0].isFatPointer()) { |
| 5923 | assert(NewAlign == (unsigned)DL.getABITypeAlignment(Ty) && |
| 5924 | "Stack not capability-aligned?" ); |
| 5925 | } |
| 5926 | if (NewAlign > Align) { |
| 5927 | // Give the stack frame object a larger alignment if needed. |
| 5928 | if (MFI.getObjectAlignment(FI->getIndex()) < NewAlign) |
| 5929 | MFI.setObjectAlignment(FI->getIndex(), NewAlign); |
| 5930 | Align = NewAlign; |
| 5931 | } |
| 5932 | } |
| 5933 | |
| 5934 | MachineMemOperand::Flags MMOFlags = |
| 5935 | isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone; |
| 5936 | SmallVector<SDValue, 16> OutLoadChains; |
| 5937 | SmallVector<SDValue, 16> OutStoreChains; |
| 5938 | SmallVector<SDValue, 32> OutChains; |
| 5939 | unsigned NumMemOps = MemOps.size(); |
| 5940 | uint64_t SrcOff = 0, DstOff = 0; |
| 5941 | for (unsigned i = 0; i != NumMemOps; ++i) { |
| 5942 | EVT VT = MemOps[i]; |
| 5943 | unsigned VTSize = VT.getSizeInBits() / 8; |
| 5944 | SDValue Value, Store; |
| 5945 | |
| 5946 | if (VTSize > Size) { |
| 5947 | // Issuing an unaligned load / store pair that overlaps with the previous |
| 5948 | // pair. Adjust the offset accordingly. |
| 5949 | assert(i == NumMemOps-1 && i != 0); |
| 5950 | SrcOff -= VTSize - Size; |
| 5951 | DstOff -= VTSize - Size; |
| 5952 | } |
| 5953 | |
| 5954 | if (CopyFromConstant && |
| 5955 | (isZeroConstant || (VT.isInteger() && !VT.isVector()))) { |
| 5956 | // It's unlikely a store of a vector immediate can be done in a single |
| 5957 | // instruction. It would require a load from a constantpool first. |
| 5958 | // We only handle zero vectors here. |
| 5959 | // FIXME: Handle other cases where store of vector immediate is done in |
| 5960 | // a single instruction. |
| 5961 | ConstantDataArraySlice SubSlice; |
| 5962 | if (SrcOff < Slice.Length) { |
| 5963 | SubSlice = Slice; |
| 5964 | SubSlice.move(SrcOff); |
| 5965 | } else { |
| 5966 | // This is an out-of-bounds access and hence UB. Pretend we read zero. |
| 5967 | SubSlice.Array = nullptr; |
| 5968 | SubSlice.Offset = 0; |
| 5969 | SubSlice.Length = VTSize; |
| 5970 | } |
| 5971 | Value = getMemsetStringVal(VT, dl, DAG, TLI, SubSlice); |
| 5972 | if (Value.getNode()) { |
| 5973 | Store = DAG.getStore(Chain, dl, Value, |
| 5974 | DAG.getMemBasePlusOffset(Dst, DstOff, dl), |
| 5975 | DstPtrInfo.getWithOffset(DstOff), Align, |
| 5976 | MMOFlags); |
| 5977 | OutChains.push_back(Store); |
| 5978 | } |
| 5979 | } |
| 5980 | |
| 5981 | if (!Store.getNode()) { |
| 5982 | // The type might not be legal for the target. This should only happen |
| 5983 | // if the type is smaller than a legal type, as on PPC, so the right |
| 5984 | // thing to do is generate a LoadExt/StoreTrunc pair. These simplify |
| 5985 | // to Load/Store if NVT==VT. |
| 5986 | // FIXME does the case above also need this? |
| 5987 | EVT NVT = TLI.getTypeToTransformTo(C, VT); |
| 5988 | assert(NVT.bitsGE(VT)); |
| 5989 | |
| 5990 | bool isDereferenceable = |
| 5991 | SrcPtrInfo.getWithOffset(SrcOff).isDereferenceable(VTSize, C, DL); |
| 5992 | MachineMemOperand::Flags SrcMMOFlags = MMOFlags; |
| 5993 | if (isDereferenceable) |
| 5994 | SrcMMOFlags |= MachineMemOperand::MODereferenceable; |
| 5995 | |
| 5996 | Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain, |
| 5997 | DAG.getMemBasePlusOffset(Src, SrcOff, dl), |
| 5998 | SrcPtrInfo.getWithOffset(SrcOff), VT, |
| 5999 | MinAlign(SrcAlign, SrcOff), SrcMMOFlags); |
| 6000 | OutLoadChains.push_back(Value.getValue(1)); |
| 6001 | |
| 6002 | Store = DAG.getTruncStore( |
| 6003 | Chain, dl, Value, DAG.getMemBasePlusOffset(Dst, DstOff, dl), |
| 6004 | DstPtrInfo.getWithOffset(DstOff), VT, Align, MMOFlags); |
| 6005 | OutStoreChains.push_back(Store); |
| 6006 | } |
| 6007 | SrcOff += VTSize; |
| 6008 | DstOff += VTSize; |
| 6009 | Size -= VTSize; |
| 6010 | } |
| 6011 | |
| 6012 | unsigned GluedLdStLimit = MaxLdStGlue == 0 ? |
| 6013 | TLI.getMaxGluedStoresPerMemcpy() : MaxLdStGlue; |
| 6014 | unsigned NumLdStInMemcpy = OutStoreChains.size(); |
| 6015 | |
| 6016 | if (NumLdStInMemcpy) { |
| 6017 | // It may be that memcpy might be converted to memset if it's memcpy |
| 6018 | // of constants. In such a case, we won't have loads and stores, but |
| 6019 | // just stores. In the absence of loads, there is nothing to gang up. |
| 6020 | if ((GluedLdStLimit <= 1) || !EnableMemCpyDAGOpt) { |
| 6021 | // If target does not care, just leave as it. |
| 6022 | for (unsigned i = 0; i < NumLdStInMemcpy; ++i) { |
| 6023 | OutChains.push_back(OutLoadChains[i]); |
| 6024 | OutChains.push_back(OutStoreChains[i]); |
| 6025 | } |
| 6026 | } else { |
| 6027 | // Ld/St less than/equal limit set by target. |
| 6028 | if (NumLdStInMemcpy <= GluedLdStLimit) { |
| 6029 | chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, 0, |
| 6030 | NumLdStInMemcpy, OutLoadChains, |
| 6031 | OutStoreChains); |
| 6032 | } else { |
| 6033 | unsigned NumberLdChain = NumLdStInMemcpy / GluedLdStLimit; |
| 6034 | unsigned RemainingLdStInMemcpy = NumLdStInMemcpy % GluedLdStLimit; |
| 6035 | unsigned GlueIter = 0; |
| 6036 | |
| 6037 | for (unsigned cnt = 0; cnt < NumberLdChain; ++cnt) { |
| 6038 | unsigned IndexFrom = NumLdStInMemcpy - GlueIter - GluedLdStLimit; |
| 6039 | unsigned IndexTo = NumLdStInMemcpy - GlueIter; |
| 6040 | |
| 6041 | chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, IndexFrom, IndexTo, |
| 6042 | OutLoadChains, OutStoreChains); |
| 6043 | GlueIter += GluedLdStLimit; |
| 6044 | } |
| 6045 | |
| 6046 | // Residual ld/st. |
| 6047 | if (RemainingLdStInMemcpy) { |
| 6048 | chainLoadsAndStoresForMemcpy(DAG, dl, OutChains, 0, |
| 6049 | RemainingLdStInMemcpy, OutLoadChains, |
| 6050 | OutStoreChains); |
| 6051 | } |
| 6052 | } |
| 6053 | } |
| 6054 | } |
| 6055 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); |
| 6056 | } |
| 6057 | |
| 6058 | static SDValue getMemmoveLoadsAndStores( |
| 6059 | SelectionDAG &DAG, const SDLoc &dl, SDValue Chain, SDValue Dst, SDValue Src, |
| 6060 | uint64_t Size, unsigned Align, bool isVol, bool AlwaysInline, |
| 6061 | bool MustPreserveCheriCapabilities, MachinePointerInfo DstPtrInfo, |
| 6062 | MachinePointerInfo SrcPtrInfo, StringRef MoveTy, |
| 6063 | CodeGenOpt::Level OptLevel) { |
| 6064 | // Turn a memmove of undef to nop. |
| 6065 | if (Src.isUndef()) |
| 6066 | return Chain; |
| 6067 | |
| 6068 | // Expand memmove to a series of load and store ops if the size operand falls |
| 6069 | // below a certain threshold. |
| 6070 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
| 6071 | const DataLayout &DL = DAG.getDataLayout(); |
| 6072 | LLVMContext &C = *DAG.getContext(); |
| 6073 | std::vector<EVT> MemOps; |
| 6074 | bool DstAlignCanChange = false; |
| 6075 | MachineFunction &MF = DAG.getMachineFunction(); |
| 6076 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 6077 | bool OptSize = shouldLowerMemFuncForSize(MF); |
| 6078 | FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst); |
| 6079 | if (FI && !MFI.isFixedObjectIndex(FI->getIndex())) |
| 6080 | DstAlignCanChange = true; |
| 6081 | unsigned SrcAlign = DAG.InferPtrAlignment(Src); |
| 6082 | if (Align > SrcAlign) |
| 6083 | SrcAlign = Align; |
| 6084 | unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemmove(OptSize); |
| 6085 | const bool FoundLowering = TLI.findOptimalMemOpLowering(MemOps, Limit, Size, |
| 6086 | (DstAlignCanChange ? 0 : Align), SrcAlign, |
| 6087 | false, false, false, false, MustPreserveCheriCapabilities, |
| 6088 | DstPtrInfo.getAddrSpace(), |
| 6089 | SrcPtrInfo.getAddrSpace(), |
| 6090 | MF.getFunction().getAttributes()); |
| 6091 | |
| 6092 | // Don't warn about inefficient memcpy if we reached the inline memmove limit |
| 6093 | // Also don't warn about copies of less than CapSize |
| 6094 | // TODO: the frontend probably shouldn't emit must-preserve-tags for such |
| 6095 | // small memcpys |
| 6096 | auto CapTy = TLI.cheriCapabilityType(); |
| 6097 | const uint64_t CapSize = CapTy.isValid() ? CapTy.getStoreSize() : 0; |
| 6098 | bool ReachedLimit = (CapSize * Limit) < Size; |
| 6099 | if (MustPreserveCheriCapabilities && !ReachedLimit && Size >= CapSize && |
| 6100 | (!FoundLowering || !MemOps[0].isFatPointer())) { |
| 6101 | LLVM_DEBUG( |
| 6102 | dbgs() << __func__ |
| 6103 | << " memmove must preserve tags but value is not statically " |
| 6104 | "known to be sufficiently aligned -> using memmove() call\n" ); |
| 6105 | if (AlwaysInline) { |
| 6106 | report_fatal_error("MustPreserveCheriCapabilities and AlwaysInline set " |
| 6107 | "but operation cannot be lowered to loads+stores!" ); |
| 6108 | } |
| 6109 | diagnoseInefficientCheriMemOp(DAG, dl.getDebugLoc(), "memmove" , OptLevel, |
| 6110 | MoveTy.empty() ? "<unknown type>" : MoveTy, |
| 6111 | std::max(1u, std::min(Align, SrcAlign)), |
| 6112 | Size, CapSize); |
| 6113 | return SDValue(); |
| 6114 | } |
| 6115 | if (!FoundLowering) |
| 6116 | return SDValue(); |
| 6117 | |
| 6118 | if (DstAlignCanChange) { |
| 6119 | Type *Ty = MemOps[0].getTypeForEVT(C); |
| 6120 | unsigned NewAlign = (unsigned)DL.getABITypeAlignment(Ty); |
| 6121 | if (MemOps[0].isFatPointer()) { |
| 6122 | assert(!DL.exceedsNaturalStackAlignment(NewAlign) && |
| 6123 | "Stack not capability-aligned?" ); |
| 6124 | } |
| 6125 | if (NewAlign > Align) { |
| 6126 | // Give the stack frame object a larger alignment if needed. |
| 6127 | if (MFI.getObjectAlignment(FI->getIndex()) < NewAlign) |
| 6128 | MFI.setObjectAlignment(FI->getIndex(), NewAlign); |
| 6129 | Align = NewAlign; |
| 6130 | } |
| 6131 | } |
| 6132 | |
| 6133 | MachineMemOperand::Flags MMOFlags = |
| 6134 | isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone; |
| 6135 | uint64_t SrcOff = 0, DstOff = 0; |
| 6136 | SmallVector<SDValue, 8> LoadValues; |
| 6137 | SmallVector<SDValue, 8> LoadChains; |
| 6138 | SmallVector<SDValue, 8> OutChains; |
| 6139 | unsigned NumMemOps = MemOps.size(); |
| 6140 | for (unsigned i = 0; i < NumMemOps; i++) { |
| 6141 | EVT VT = MemOps[i]; |
| 6142 | unsigned VTSize = VT.getSizeInBits() / 8; |
| 6143 | SDValue Value; |
| 6144 | |
| 6145 | bool isDereferenceable = |
| 6146 | SrcPtrInfo.getWithOffset(SrcOff).isDereferenceable(VTSize, C, DL); |
| 6147 | MachineMemOperand::Flags SrcMMOFlags = MMOFlags; |
| 6148 | if (isDereferenceable) |
| 6149 | SrcMMOFlags |= MachineMemOperand::MODereferenceable; |
| 6150 | |
| 6151 | Value = |
| 6152 | DAG.getLoad(VT, dl, Chain, DAG.getMemBasePlusOffset(Src, SrcOff, dl), |
| 6153 | SrcPtrInfo.getWithOffset(SrcOff), SrcAlign, SrcMMOFlags); |
| 6154 | LoadValues.push_back(Value); |
| 6155 | LoadChains.push_back(Value.getValue(1)); |
| 6156 | SrcOff += VTSize; |
| 6157 | } |
| 6158 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains); |
| 6159 | OutChains.clear(); |
| 6160 | for (unsigned i = 0; i < NumMemOps; i++) { |
| 6161 | EVT VT = MemOps[i]; |
| 6162 | unsigned VTSize = VT.getSizeInBits() / 8; |
| 6163 | SDValue Store; |
| 6164 | |
| 6165 | Store = DAG.getStore(Chain, dl, LoadValues[i], |
| 6166 | DAG.getMemBasePlusOffset(Dst, DstOff, dl), |
| 6167 | DstPtrInfo.getWithOffset(DstOff), Align, MMOFlags); |
| 6168 | OutChains.push_back(Store); |
| 6169 | DstOff += VTSize; |
| 6170 | } |
| 6171 | |
| 6172 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); |
| 6173 | } |
| 6174 | |
| 6175 | /// Lower the call to 'memset' intrinsic function into a series of store |
| 6176 | /// operations. |
| 6177 | /// |
| 6178 | /// \param DAG Selection DAG where lowered code is placed. |
| 6179 | /// \param dl Link to corresponding IR location. |
| 6180 | /// \param Chain Control flow dependency. |
| 6181 | /// \param Dst Pointer to destination memory location. |
| 6182 | /// \param Src Value of byte to write into the memory. |
| 6183 | /// \param Size Number of bytes to write. |
| 6184 | /// \param Align Alignment of the destination in bytes. |
| 6185 | /// \param isVol True if destination is volatile. |
| 6186 | /// \param DstPtrInfo IR information on the memory pointer. |
| 6187 | /// \returns New head in the control flow, if lowering was successful, empty |
| 6188 | /// SDValue otherwise. |
| 6189 | /// |
| 6190 | /// The function tries to replace 'llvm.memset' intrinsic with several store |
| 6191 | /// operations and value calculation code. This is usually profitable for small |
| 6192 | /// memory size. |
| 6193 | static SDValue getMemsetStores(SelectionDAG &DAG, const SDLoc &dl, |
| 6194 | SDValue Chain, SDValue Dst, SDValue Src, |
| 6195 | uint64_t Size, unsigned Align, bool isVol, |
| 6196 | MachinePointerInfo DstPtrInfo) { |
| 6197 | // Turn a memset of undef to nop. |
| 6198 | if (Src.isUndef()) |
| 6199 | return Chain; |
| 6200 | |
| 6201 | // Expand memset to a series of load/store ops if the size operand |
| 6202 | // falls below a certain threshold. |
| 6203 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
| 6204 | std::vector<EVT> MemOps; |
| 6205 | bool DstAlignCanChange = false; |
| 6206 | MachineFunction &MF = DAG.getMachineFunction(); |
| 6207 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 6208 | bool OptSize = shouldLowerMemFuncForSize(MF); |
| 6209 | FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst); |
| 6210 | if (FI && !MFI.isFixedObjectIndex(FI->getIndex())) |
| 6211 | DstAlignCanChange = true; |
| 6212 | bool IsZeroVal = |
| 6213 | isa<ConstantSDNode>(Src) && cast<ConstantSDNode>(Src)->isNullValue(); |
| 6214 | if (!TLI.findOptimalMemOpLowering(MemOps, TLI.getMaxStoresPerMemset(OptSize), |
| 6215 | Size, (DstAlignCanChange ? 0 : Align), 0, |
| 6216 | true, IsZeroVal, false, true, false, |
| 6217 | DstPtrInfo.getAddrSpace(), ~0u, |
| 6218 | MF.getFunction().getAttributes())) |
| 6219 | return SDValue(); |
| 6220 | |
| 6221 | if (DstAlignCanChange) { |
| 6222 | Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext()); |
| 6223 | unsigned NewAlign = (unsigned)DAG.getDataLayout().getABITypeAlignment(Ty); |
| 6224 | if (MemOps[0].isFatPointer()) { |
| 6225 | assert(!DAG.getDataLayout().exceedsNaturalStackAlignment(NewAlign) && |
| 6226 | "Stack not capability-aligned?" ); |
| 6227 | } |
| 6228 | if (NewAlign > Align) { |
| 6229 | // Give the stack frame object a larger alignment if needed. |
| 6230 | if (MFI.getObjectAlignment(FI->getIndex()) < NewAlign) |
| 6231 | MFI.setObjectAlignment(FI->getIndex(), NewAlign); |
| 6232 | Align = NewAlign; |
| 6233 | } |
| 6234 | } |
| 6235 | |
| 6236 | SmallVector<SDValue, 8> OutChains; |
| 6237 | uint64_t DstOff = 0; |
| 6238 | unsigned NumMemOps = MemOps.size(); |
| 6239 | |
| 6240 | // Find the largest store and generate the bit pattern for it. |
| 6241 | EVT LargestVT = MemOps[0]; |
| 6242 | for (unsigned i = 1; i < NumMemOps; i++) |
| 6243 | if (MemOps[i].bitsGT(LargestVT)) |
| 6244 | LargestVT = MemOps[i]; |
| 6245 | SDValue MemSetValue = getMemsetValue(Src, LargestVT, DAG, dl); |
| 6246 | |
| 6247 | for (unsigned i = 0; i < NumMemOps; i++) { |
| 6248 | EVT VT = MemOps[i]; |
| 6249 | unsigned VTSize = VT.getSizeInBits() / 8; |
| 6250 | if (VTSize > Size) { |
| 6251 | // Issuing an unaligned load / store pair that overlaps with the previous |
| 6252 | // pair. Adjust the offset accordingly. |
| 6253 | assert(i == NumMemOps-1 && i != 0); |
| 6254 | DstOff -= VTSize - Size; |
| 6255 | } |
| 6256 | |
| 6257 | // If this store is smaller than the largest store see whether we can get |
| 6258 | // the smaller value for free with a truncate. |
| 6259 | SDValue Value = MemSetValue; |
| 6260 | if (VT.bitsLT(LargestVT)) { |
| 6261 | if (!LargestVT.isVector() && !VT.isVector() && |
| 6262 | TLI.isTruncateFree(LargestVT, VT)) |
| 6263 | Value = DAG.getNode(ISD::TRUNCATE, dl, VT, MemSetValue); |
| 6264 | else |
| 6265 | Value = getMemsetValue(Src, VT, DAG, dl); |
| 6266 | } |
| 6267 | assert(Value.getValueType() == VT && "Value with wrong type." ); |
| 6268 | SDValue Store = DAG.getStore( |
| 6269 | Chain, dl, Value, DAG.getMemBasePlusOffset(Dst, DstOff, dl), |
| 6270 | DstPtrInfo.getWithOffset(DstOff), Align, |
| 6271 | isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone); |
| 6272 | OutChains.push_back(Store); |
| 6273 | DstOff += VT.getSizeInBits() / 8; |
| 6274 | Size -= VTSize; |
| 6275 | } |
| 6276 | |
| 6277 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); |
| 6278 | } |
| 6279 | |
| 6280 | SDValue SelectionDAG::getPointerAdd(const SDLoc dl, SDValue Ptr, SDValue Offset, |
| 6281 | const SDNodeFlags Flags) { |
| 6282 | assert(Offset.getValueType().isInteger()); |
| 6283 | EVT BasePtrVT = Ptr.getValueType(); |
| 6284 | if (BasePtrVT.isFatPointer()) { |
| 6285 | if (auto *Constant = dyn_cast<ConstantSDNode>(Offset.getNode())) { |
| 6286 | if (Constant->isNullValue()) |
| 6287 | return Ptr; |
| 6288 | } |
| 6289 | return getNode(ISD::PTRADD, dl, BasePtrVT, Ptr, Offset, Flags); |
| 6290 | } |
| 6291 | return getNode(ISD::ADD, dl, BasePtrVT, Ptr, Offset, Flags); |
| 6292 | } |
| 6293 | |
| 6294 | SDValue SelectionDAG::getCSetBounds(SDValue Val, SDValue Length, |
| 6295 | bool CSetBoundsStatsAlreadyLogged) { |
| 6296 | if (!CSetBoundsStatsAlreadyLogged && cheri::ShouldCollectCSetBoundsStats) { |
| 6297 | Optional<uint64_t> SizeConst; |
| 6298 | if (ConstantSDNode *Constant = dyn_cast<ConstantSDNode>(Length.getNode())) { |
| 6299 | SizeConst = Constant->getZExtValue(); |
| 6300 | } |
| 6301 | cheri::CSetBoundsStats->add( |
| 6302 | 1, SizeConst, "SelectionDAG::getCSetBounds" , |
| 6303 | cheri::SetBoundsPointerSource::Unknown, "<unnkonwn reason>" , |
| 6304 | cheri::inferSourceLocation(Val.getDebugLoc(), |
| 6305 | getMachineFunction().getName())); |
| 6306 | } |
| 6307 | SDLoc DL(Val); |
| 6308 | Intrinsic::ID SetBounds = Intrinsic::cheri_cap_bounds_set; |
| 6309 | // Using the bounded stack cap intrinisic allows reuse of the same register: |
| 6310 | if (isa<FrameIndexSDNode>(Val.getNode())) |
| 6311 | SetBounds = Intrinsic::cheri_bounded_stack_cap; |
| 6312 | MVT SizeVT = MVT::getIntegerVT(getDataLayout().getPointerSizeInBits(0)); |
| 6313 | return getNode(ISD::INTRINSIC_WO_CHAIN, DL, Val.getValueType(), |
| 6314 | getConstant(SetBounds, DL, SizeVT), Val, Length); |
| 6315 | } |
| 6316 | |
| 6317 | static void checkAddrSpaceIsValidForLibcall(const TargetLowering *TLI, |
| 6318 | unsigned AS) { |
| 6319 | // Lowering memcpy / memset / memmove intrinsics to calls is only valid if all |
| 6320 | // pointer operands can be losslessly bitcasted to pointers of address space 0 |
| 6321 | if (AS != 0 && !TLI->isNoopAddrSpaceCast(AS, 0)) { |
| 6322 | report_fatal_error("cannot lower memory intrinsic in address space " + |
| 6323 | Twine(AS)); |
| 6324 | } |
| 6325 | } |
| 6326 | |
| 6327 | SDValue SelectionDAG::getMemcpy(SDValue Chain, const SDLoc &dl, SDValue Dst, |
| 6328 | SDValue Src, SDValue Size, unsigned Align, |
| 6329 | bool isVol, bool AlwaysInline, bool isTailCall, |
| 6330 | bool MustPreserveCheriCapabilities, |
| 6331 | MachinePointerInfo DstPtrInfo, |
| 6332 | MachinePointerInfo SrcPtrInfo, |
| 6333 | StringRef CopyType) { |
| 6334 | assert(Align && "The SDAG layer expects explicit alignment and reserves 0" ); |
| 6335 | LLVM_DEBUG(dbgs() << "DAG.getMemcpy() align=" << Align << " size=" ; |
| 6336 | Size.dump();); |
| 6337 | |
| 6338 | // Check to see if we should lower the memcpy to loads and stores first. |
| 6339 | // For cases within the target-specified limits, this is the best choice. |
| 6340 | ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); |
| 6341 | if (MustPreserveCheriCapabilities) |
| 6342 | assert(TLI->cheriCapabilityType().isValid()); |
| 6343 | if (ConstantSize) { |
| 6344 | // Memcpy with size zero? Just return the original chain. |
| 6345 | if (ConstantSize->isNullValue()) |
| 6346 | return Chain; |
| 6347 | |
| 6348 | SDValue Result; |
| 6349 | Result = getMemcpyLoadsAndStores( |
| 6350 | *this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), Align, isVol, |
| 6351 | false, MustPreserveCheriCapabilities, DstPtrInfo, SrcPtrInfo, CopyType, |
| 6352 | OptLevel); |
| 6353 | if (Result.getNode()) |
| 6354 | return Result; |
| 6355 | } |
| 6356 | |
| 6357 | // Then check to see if we should lower the memcpy with target-specific |
| 6358 | // code. If the target chooses to do this, this is the next best. |
| 6359 | if (TSI) { |
| 6360 | SDValue Result = TSI->EmitTargetCodeForMemcpy( |
| 6361 | *this, dl, Chain, Dst, Src, Size, Align, isVol, AlwaysInline, |
| 6362 | MustPreserveCheriCapabilities, DstPtrInfo, SrcPtrInfo); |
| 6363 | if (Result.getNode()) |
| 6364 | return Result; |
| 6365 | } |
| 6366 | |
| 6367 | // If we really need inline code and the target declined to provide it, |
| 6368 | // use a (potentially long) sequence of loads and stores. |
| 6369 | if (AlwaysInline) { |
| 6370 | assert(ConstantSize && "AlwaysInline requires a constant size!" ); |
| 6371 | return getMemcpyLoadsAndStores( |
| 6372 | *this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), Align, isVol, |
| 6373 | true, MustPreserveCheriCapabilities, DstPtrInfo, SrcPtrInfo, CopyType, |
| 6374 | OptLevel); |
| 6375 | } |
| 6376 | |
| 6377 | checkAddrSpaceIsValidForLibcall(TLI, DstPtrInfo.getAddrSpace()); |
| 6378 | checkAddrSpaceIsValidForLibcall(TLI, SrcPtrInfo.getAddrSpace()); |
| 6379 | |
| 6380 | // FIXME: If the memcpy is volatile (isVol), lowering it to a plain libc |
| 6381 | // memcpy is not guaranteed to be safe. libc memcpys aren't required to |
| 6382 | // respect volatile, so they may do things like read or write memory |
| 6383 | // beyond the given memory regions. But fixing this isn't easy, and most |
| 6384 | // people don't care. |
| 6385 | |
| 6386 | // Emit a library call. |
| 6387 | TargetLowering::ArgListTy Args; |
| 6388 | TargetLowering::ArgListEntry Entry; |
| 6389 | Entry.Ty = Dst.getValueType().getTypeForEVT(*getContext()); |
| 6390 | Entry.Node = Dst; Args.push_back(Entry); |
| 6391 | Entry.Ty = Src.getValueType().getTypeForEVT(*getContext()); |
| 6392 | Entry.Node = Src; Args.push_back(Entry); |
| 6393 | |
| 6394 | Entry.Ty = getDataLayout().getIntPtrType(*getContext()); |
| 6395 | Entry.Node = Size; Args.push_back(Entry); |
| 6396 | // FIXME: pass in SDLoc |
| 6397 | TargetLowering::CallLoweringInfo CLI(*this); |
| 6398 | CLI.setDebugLoc(dl) |
| 6399 | .setChain(Chain) |
| 6400 | .setLibCallee( |
| 6401 | TLI->getLibcallCallingConv(RTLIB::MEMCPY), |
| 6402 | Dst.getValueType().getTypeForEVT(*getContext()), |
| 6403 | getExternalFunctionSymbol(TLI->getLibcallName(RTLIB::MEMCPY)), |
| 6404 | std::move(Args)) |
| 6405 | .setDiscardResult() |
| 6406 | .setTailCall(isTailCall); |
| 6407 | |
| 6408 | std::pair<SDValue,SDValue> CallResult = TLI->LowerCallTo(CLI); |
| 6409 | return CallResult.second; |
| 6410 | } |
| 6411 | |
| 6412 | SDValue SelectionDAG::getAtomicMemcpy(SDValue Chain, const SDLoc &dl, |
| 6413 | SDValue Dst, unsigned DstAlign, |
| 6414 | SDValue Src, unsigned SrcAlign, |
| 6415 | SDValue Size, Type *SizeTy, |
| 6416 | unsigned ElemSz, bool isTailCall, |
| 6417 | MachinePointerInfo DstPtrInfo, |
| 6418 | MachinePointerInfo SrcPtrInfo) { |
| 6419 | // Emit a library call. |
| 6420 | TargetLowering::ArgListTy Args; |
| 6421 | TargetLowering::ArgListEntry Entry; |
| 6422 | Entry.Ty = getDataLayout().getIntPtrType(*getContext()); |
| 6423 | Entry.Node = Dst; |
| 6424 | Args.push_back(Entry); |
| 6425 | |
| 6426 | Entry.Node = Src; |
| 6427 | Args.push_back(Entry); |
| 6428 | |
| 6429 | Entry.Ty = SizeTy; |
| 6430 | Entry.Node = Size; |
| 6431 | Args.push_back(Entry); |
| 6432 | |
| 6433 | RTLIB::Libcall LibraryCall = |
| 6434 | RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(ElemSz); |
| 6435 | if (LibraryCall == RTLIB::UNKNOWN_LIBCALL) |
| 6436 | report_fatal_error("Unsupported element size" ); |
| 6437 | |
| 6438 | TargetLowering::CallLoweringInfo CLI(*this); |
| 6439 | CLI.setDebugLoc(dl) |
| 6440 | .setChain(Chain) |
| 6441 | .setLibCallee(TLI->getLibcallCallingConv(LibraryCall), |
| 6442 | Type::getVoidTy(*getContext()), |
| 6443 | getExternalFunctionSymbol(TLI->getLibcallName(LibraryCall)), |
| 6444 | std::move(Args)) |
| 6445 | .setDiscardResult() |
| 6446 | .setTailCall(isTailCall); |
| 6447 | |
| 6448 | std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI); |
| 6449 | return CallResult.second; |
| 6450 | } |
| 6451 | |
| 6452 | SDValue SelectionDAG::getMemmove(SDValue Chain, const SDLoc &dl, SDValue Dst, |
| 6453 | SDValue Src, SDValue Size, unsigned Align, |
| 6454 | bool isVol, bool isTailCall, |
| 6455 | bool MustPreserveCheriCapabilities, |
| 6456 | MachinePointerInfo DstPtrInfo, |
| 6457 | MachinePointerInfo SrcPtrInfo, |
| 6458 | StringRef MoveType) { |
| 6459 | assert(Align && "The SDAG layer expects explicit alignment and reserves 0" ); |
| 6460 | |
| 6461 | // Check to see if we should lower the memmove to loads and stores first. |
| 6462 | // For cases within the target-specified limits, this is the best choice. |
| 6463 | ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); |
| 6464 | if (ConstantSize) { |
| 6465 | // Memmove with size zero? Just return the original chain. |
| 6466 | if (ConstantSize->isNullValue()) |
| 6467 | return Chain; |
| 6468 | |
| 6469 | SDValue Result; |
| 6470 | Result = getMemmoveLoadsAndStores( |
| 6471 | *this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), Align, isVol, |
| 6472 | false, MustPreserveCheriCapabilities, DstPtrInfo, SrcPtrInfo, MoveType, |
| 6473 | OptLevel); |
| 6474 | if (Result.getNode()) |
| 6475 | return Result; |
| 6476 | } |
| 6477 | |
| 6478 | // Then check to see if we should lower the memmove with target-specific |
| 6479 | // code. If the target chooses to do this, this is the next best. |
| 6480 | if (TSI) { |
| 6481 | SDValue Result = TSI->EmitTargetCodeForMemmove( |
| 6482 | *this, dl, Chain, Dst, Src, Size, Align, isVol, |
| 6483 | MustPreserveCheriCapabilities, DstPtrInfo, SrcPtrInfo); |
| 6484 | if (Result.getNode()) |
| 6485 | return Result; |
| 6486 | } |
| 6487 | |
| 6488 | checkAddrSpaceIsValidForLibcall(TLI, DstPtrInfo.getAddrSpace()); |
| 6489 | checkAddrSpaceIsValidForLibcall(TLI, SrcPtrInfo.getAddrSpace()); |
| 6490 | |
| 6491 | // FIXME: If the memmove is volatile, lowering it to plain libc memmove may |
| 6492 | // not be safe. See memcpy above for more details. |
| 6493 | |
| 6494 | // Emit a library call. |
| 6495 | TargetLowering::ArgListTy Args; |
| 6496 | TargetLowering::ArgListEntry Entry; |
| 6497 | Entry.Ty = Dst.getValueType().getTypeForEVT(*getContext()); |
| 6498 | Entry.Node = Dst; Args.push_back(Entry); |
| 6499 | Entry.Ty = Src.getValueType().getTypeForEVT(*getContext()); |
| 6500 | Entry.Node = Src; Args.push_back(Entry); |
| 6501 | |
| 6502 | Entry.Ty = getDataLayout().getIntPtrType(*getContext()); |
| 6503 | Entry.Node = Size; Args.push_back(Entry); |
| 6504 | // FIXME: pass in SDLoc |
| 6505 | TargetLowering::CallLoweringInfo CLI(*this); |
| 6506 | CLI.setDebugLoc(dl) |
| 6507 | .setChain(Chain) |
| 6508 | .setLibCallee( |
| 6509 | TLI->getLibcallCallingConv(RTLIB::MEMMOVE), |
| 6510 | Dst.getValueType().getTypeForEVT(*getContext()), |
| 6511 | getExternalFunctionSymbol(TLI->getLibcallName(RTLIB::MEMMOVE)), |
| 6512 | std::move(Args)) |
| 6513 | .setDiscardResult() |
| 6514 | .setTailCall(isTailCall); |
| 6515 | |
| 6516 | std::pair<SDValue,SDValue> CallResult = TLI->LowerCallTo(CLI); |
| 6517 | return CallResult.second; |
| 6518 | } |
| 6519 | |
| 6520 | SDValue SelectionDAG::getAtomicMemmove(SDValue Chain, const SDLoc &dl, |
| 6521 | SDValue Dst, unsigned DstAlign, |
| 6522 | SDValue Src, unsigned SrcAlign, |
| 6523 | SDValue Size, Type *SizeTy, |
| 6524 | unsigned ElemSz, bool isTailCall, |
| 6525 | MachinePointerInfo DstPtrInfo, |
| 6526 | MachinePointerInfo SrcPtrInfo) { |
| 6527 | // Emit a library call. |
| 6528 | TargetLowering::ArgListTy Args; |
| 6529 | TargetLowering::ArgListEntry Entry; |
| 6530 | Entry.Ty = getDataLayout().getIntPtrType(*getContext()); |
| 6531 | Entry.Node = Dst; |
| 6532 | Args.push_back(Entry); |
| 6533 | |
| 6534 | Entry.Node = Src; |
| 6535 | Args.push_back(Entry); |
| 6536 | |
| 6537 | Entry.Ty = SizeTy; |
| 6538 | Entry.Node = Size; |
| 6539 | Args.push_back(Entry); |
| 6540 | |
| 6541 | RTLIB::Libcall LibraryCall = |
| 6542 | RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(ElemSz); |
| 6543 | if (LibraryCall == RTLIB::UNKNOWN_LIBCALL) |
| 6544 | report_fatal_error("Unsupported element size" ); |
| 6545 | |
| 6546 | TargetLowering::CallLoweringInfo CLI(*this); |
| 6547 | CLI.setDebugLoc(dl) |
| 6548 | .setChain(Chain) |
| 6549 | .setLibCallee(TLI->getLibcallCallingConv(LibraryCall), |
| 6550 | Type::getVoidTy(*getContext()), |
| 6551 | getExternalFunctionSymbol(TLI->getLibcallName(LibraryCall)), |
| 6552 | std::move(Args)) |
| 6553 | .setDiscardResult() |
| 6554 | .setTailCall(isTailCall); |
| 6555 | |
| 6556 | std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI); |
| 6557 | return CallResult.second; |
| 6558 | } |
| 6559 | |
| 6560 | SDValue SelectionDAG::getMemset(SDValue Chain, const SDLoc &dl, SDValue Dst, |
| 6561 | SDValue Src, SDValue Size, unsigned Align, |
| 6562 | bool isVol, bool isTailCall, |
| 6563 | MachinePointerInfo DstPtrInfo) { |
| 6564 | assert(Align && "The SDAG layer expects explicit alignment and reserves 0" ); |
| 6565 | |
| 6566 | // Check to see if we should lower the memset to stores first. |
| 6567 | // For cases within the target-specified limits, this is the best choice. |
| 6568 | ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); |
| 6569 | if (ConstantSize) { |
| 6570 | // Memset with size zero? Just return the original chain. |
| 6571 | if (ConstantSize->isNullValue()) |
| 6572 | return Chain; |
| 6573 | |
| 6574 | SDValue Result = |
| 6575 | getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), |
| 6576 | Align, isVol, DstPtrInfo); |
| 6577 | |
| 6578 | if (Result.getNode()) |
| 6579 | return Result; |
| 6580 | } |
| 6581 | |
| 6582 | // Then check to see if we should lower the memset with target-specific |
| 6583 | // code. If the target chooses to do this, this is the next best. |
| 6584 | if (TSI) { |
| 6585 | SDValue Result = TSI->EmitTargetCodeForMemset( |
| 6586 | *this, dl, Chain, Dst, Src, Size, Align, isVol, DstPtrInfo); |
| 6587 | if (Result.getNode()) |
| 6588 | return Result; |
| 6589 | } |
| 6590 | |
| 6591 | checkAddrSpaceIsValidForLibcall(TLI, DstPtrInfo.getAddrSpace()); |
| 6592 | |
| 6593 | // Emit a library call. |
| 6594 | TargetLowering::ArgListTy Args; |
| 6595 | TargetLowering::ArgListEntry Entry; |
| 6596 | Entry.Node = Dst; Entry.Ty = Dst.getValueType().getTypeForEVT(*getContext()); |
| 6597 | Args.push_back(Entry); |
| 6598 | Entry.Node = Src; |
| 6599 | Entry.Ty = Src.getValueType().getTypeForEVT(*getContext()); |
| 6600 | Args.push_back(Entry); |
| 6601 | Entry.Node = Size; |
| 6602 | Entry.Ty = getDataLayout().getIntPtrType(*getContext()); |
| 6603 | Args.push_back(Entry); |
| 6604 | |
| 6605 | // FIXME: pass in SDLoc |
| 6606 | TargetLowering::CallLoweringInfo CLI(*this); |
| 6607 | CLI.setDebugLoc(dl) |
| 6608 | .setChain(Chain) |
| 6609 | .setLibCallee( |
| 6610 | TLI->getLibcallCallingConv(RTLIB::MEMSET), |
| 6611 | Dst.getValueType().getTypeForEVT(*getContext()), |
| 6612 | getExternalFunctionSymbol(TLI->getLibcallName(RTLIB::MEMSET)), |
| 6613 | std::move(Args)) |
| 6614 | .setDiscardResult() |
| 6615 | .setTailCall(isTailCall); |
| 6616 | |
| 6617 | std::pair<SDValue,SDValue> CallResult = TLI->LowerCallTo(CLI); |
| 6618 | return CallResult.second; |
| 6619 | } |
| 6620 | |
| 6621 | SDValue SelectionDAG::getAtomicMemset(SDValue Chain, const SDLoc &dl, |
| 6622 | SDValue Dst, unsigned DstAlign, |
| 6623 | SDValue Value, SDValue Size, Type *SizeTy, |
| 6624 | unsigned ElemSz, bool isTailCall, |
| 6625 | MachinePointerInfo DstPtrInfo) { |
| 6626 | // Emit a library call. |
| 6627 | TargetLowering::ArgListTy Args; |
| 6628 | TargetLowering::ArgListEntry Entry; |
| 6629 | Entry.Ty = getDataLayout().getIntPtrType(*getContext()); |
| 6630 | Entry.Node = Dst; |
| 6631 | Args.push_back(Entry); |
| 6632 | |
| 6633 | Entry.Ty = Type::getInt8Ty(*getContext()); |
| 6634 | Entry.Node = Value; |
| 6635 | Args.push_back(Entry); |
| 6636 | |
| 6637 | Entry.Ty = SizeTy; |
| 6638 | Entry.Node = Size; |
| 6639 | Args.push_back(Entry); |
| 6640 | |
| 6641 | RTLIB::Libcall LibraryCall = |
| 6642 | RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(ElemSz); |
| 6643 | if (LibraryCall == RTLIB::UNKNOWN_LIBCALL) |
| 6644 | report_fatal_error("Unsupported element size" ); |
| 6645 | |
| 6646 | TargetLowering::CallLoweringInfo CLI(*this); |
| 6647 | CLI.setDebugLoc(dl) |
| 6648 | .setChain(Chain) |
| 6649 | .setLibCallee(TLI->getLibcallCallingConv(LibraryCall), |
| 6650 | Type::getVoidTy(*getContext()), |
| 6651 | getExternalFunctionSymbol(TLI->getLibcallName(LibraryCall)), |
| 6652 | std::move(Args)) |
| 6653 | .setDiscardResult() |
| 6654 | .setTailCall(isTailCall); |
| 6655 | |
| 6656 | std::pair<SDValue, SDValue> CallResult = TLI->LowerCallTo(CLI); |
| 6657 | return CallResult.second; |
| 6658 | } |
| 6659 | |
| 6660 | SDValue SelectionDAG::getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT, |
| 6661 | SDVTList VTList, ArrayRef<SDValue> Ops, |
| 6662 | MachineMemOperand *MMO) { |
| 6663 | FoldingSetNodeID ID; |
| 6664 | ID.AddInteger(MemVT.getRawBits()); |
| 6665 | AddNodeIDNode(ID, Opcode, VTList, Ops); |
| 6666 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 6667 | void* IP = nullptr; |
| 6668 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 6669 | cast<AtomicSDNode>(E)->refineAlignment(MMO); |
| 6670 | return SDValue(E, 0); |
| 6671 | } |
| 6672 | |
| 6673 | auto *N = newSDNode<AtomicSDNode>(Opcode, dl.getIROrder(), dl.getDebugLoc(), |
| 6674 | VTList, MemVT, MMO); |
| 6675 | createOperands(N, Ops); |
| 6676 | |
| 6677 | CSEMap.InsertNode(N, IP); |
| 6678 | InsertNode(N); |
| 6679 | return SDValue(N, 0); |
| 6680 | } |
| 6681 | |
| 6682 | SDValue SelectionDAG::getAtomicCmpSwap(unsigned Opcode, const SDLoc &dl, |
| 6683 | EVT MemVT, SDVTList VTs, SDValue Chain, |
| 6684 | SDValue Ptr, SDValue Cmp, SDValue Swp, |
| 6685 | MachineMemOperand *MMO) { |
| 6686 | assert(Opcode == ISD::ATOMIC_CMP_SWAP || |
| 6687 | Opcode == ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS); |
| 6688 | assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types" ); |
| 6689 | |
| 6690 | SDValue Ops[] = {Chain, Ptr, Cmp, Swp}; |
| 6691 | return getAtomic(Opcode, dl, MemVT, VTs, Ops, MMO); |
| 6692 | } |
| 6693 | |
| 6694 | SDValue SelectionDAG::getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT, |
| 6695 | SDValue Chain, SDValue Ptr, SDValue Val, |
| 6696 | MachineMemOperand *MMO) { |
| 6697 | assert((Opcode == ISD::ATOMIC_LOAD_ADD || |
| 6698 | Opcode == ISD::ATOMIC_LOAD_SUB || |
| 6699 | Opcode == ISD::ATOMIC_LOAD_AND || |
| 6700 | Opcode == ISD::ATOMIC_LOAD_CLR || |
| 6701 | Opcode == ISD::ATOMIC_LOAD_OR || |
| 6702 | Opcode == ISD::ATOMIC_LOAD_XOR || |
| 6703 | Opcode == ISD::ATOMIC_LOAD_NAND || |
| 6704 | Opcode == ISD::ATOMIC_LOAD_MIN || |
| 6705 | Opcode == ISD::ATOMIC_LOAD_MAX || |
| 6706 | Opcode == ISD::ATOMIC_LOAD_UMIN || |
| 6707 | Opcode == ISD::ATOMIC_LOAD_UMAX || |
| 6708 | Opcode == ISD::ATOMIC_LOAD_FADD || |
| 6709 | Opcode == ISD::ATOMIC_LOAD_FSUB || |
| 6710 | Opcode == ISD::ATOMIC_SWAP || |
| 6711 | Opcode == ISD::ATOMIC_STORE) && |
| 6712 | "Invalid Atomic Op" ); |
| 6713 | |
| 6714 | EVT VT = Val.getValueType(); |
| 6715 | |
| 6716 | SDVTList VTs = Opcode == ISD::ATOMIC_STORE ? getVTList(MVT::Other) : |
| 6717 | getVTList(VT, MVT::Other); |
| 6718 | SDValue Ops[] = {Chain, Ptr, Val}; |
| 6719 | return getAtomic(Opcode, dl, MemVT, VTs, Ops, MMO); |
| 6720 | } |
| 6721 | |
| 6722 | SDValue SelectionDAG::getAtomic(unsigned Opcode, const SDLoc &dl, EVT MemVT, |
| 6723 | EVT VT, SDValue Chain, SDValue Ptr, |
| 6724 | MachineMemOperand *MMO) { |
| 6725 | assert(Opcode == ISD::ATOMIC_LOAD && "Invalid Atomic Op" ); |
| 6726 | |
| 6727 | SDVTList VTs = getVTList(VT, MVT::Other); |
| 6728 | SDValue Ops[] = {Chain, Ptr}; |
| 6729 | return getAtomic(Opcode, dl, MemVT, VTs, Ops, MMO); |
| 6730 | } |
| 6731 | |
| 6732 | /// getMergeValues - Create a MERGE_VALUES node from the given operands. |
| 6733 | SDValue SelectionDAG::getMergeValues(ArrayRef<SDValue> Ops, const SDLoc &dl) { |
| 6734 | if (Ops.size() == 1) |
| 6735 | return Ops[0]; |
| 6736 | |
| 6737 | SmallVector<EVT, 4> VTs; |
| 6738 | VTs.reserve(Ops.size()); |
| 6739 | for (unsigned i = 0; i < Ops.size(); ++i) |
| 6740 | VTs.push_back(Ops[i].getValueType()); |
| 6741 | return getNode(ISD::MERGE_VALUES, dl, getVTList(VTs), Ops); |
| 6742 | } |
| 6743 | |
| 6744 | SDValue SelectionDAG::getMemIntrinsicNode( |
| 6745 | unsigned Opcode, const SDLoc &dl, SDVTList VTList, ArrayRef<SDValue> Ops, |
| 6746 | EVT MemVT, MachinePointerInfo PtrInfo, unsigned Align, |
| 6747 | MachineMemOperand::Flags Flags, unsigned Size) { |
| 6748 | if (Align == 0) // Ensure that codegen never sees alignment 0 |
| 6749 | Align = getEVTAlignment(MemVT); |
| 6750 | |
| 6751 | if (!Size) |
| 6752 | Size = MemVT.getStoreSize(); |
| 6753 | |
| 6754 | MachineFunction &MF = getMachineFunction(); |
| 6755 | MachineMemOperand *MMO = |
| 6756 | MF.getMachineMemOperand(PtrInfo, Flags, Size, Align); |
| 6757 | |
| 6758 | return getMemIntrinsicNode(Opcode, dl, VTList, Ops, MemVT, MMO); |
| 6759 | } |
| 6760 | |
| 6761 | SDValue SelectionDAG::getMemIntrinsicNode(unsigned Opcode, const SDLoc &dl, |
| 6762 | SDVTList VTList, |
| 6763 | ArrayRef<SDValue> Ops, EVT MemVT, |
| 6764 | MachineMemOperand *MMO) { |
| 6765 | assert((Opcode == ISD::INTRINSIC_VOID || |
| 6766 | Opcode == ISD::INTRINSIC_W_CHAIN || |
| 6767 | Opcode == ISD::PREFETCH || |
| 6768 | Opcode == ISD::LIFETIME_START || |
| 6769 | Opcode == ISD::LIFETIME_END || |
| 6770 | ((int)Opcode <= std::numeric_limits<int>::max() && |
| 6771 | (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) && |
| 6772 | "Opcode is not a memory-accessing opcode!" ); |
| 6773 | |
| 6774 | // Memoize the node unless it returns a flag. |
| 6775 | MemIntrinsicSDNode *N; |
| 6776 | if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) { |
| 6777 | FoldingSetNodeID ID; |
| 6778 | AddNodeIDNode(ID, Opcode, VTList, Ops); |
| 6779 | ID.AddInteger(getSyntheticNodeSubclassData<MemIntrinsicSDNode>( |
| 6780 | Opcode, dl.getIROrder(), VTList, MemVT, MMO)); |
| 6781 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 6782 | void *IP = nullptr; |
| 6783 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 6784 | cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO); |
| 6785 | return SDValue(E, 0); |
| 6786 | } |
| 6787 | |
| 6788 | N = newSDNode<MemIntrinsicSDNode>(Opcode, dl.getIROrder(), dl.getDebugLoc(), |
| 6789 | VTList, MemVT, MMO); |
| 6790 | createOperands(N, Ops); |
| 6791 | |
| 6792 | CSEMap.InsertNode(N, IP); |
| 6793 | } else { |
| 6794 | N = newSDNode<MemIntrinsicSDNode>(Opcode, dl.getIROrder(), dl.getDebugLoc(), |
| 6795 | VTList, MemVT, MMO); |
| 6796 | createOperands(N, Ops); |
| 6797 | } |
| 6798 | InsertNode(N); |
| 6799 | return SDValue(N, 0); |
| 6800 | } |
| 6801 | |
| 6802 | SDValue SelectionDAG::getLifetimeNode(bool IsStart, const SDLoc &dl, |
| 6803 | SDValue Chain, int FrameIndex, |
| 6804 | int64_t Size, int64_t Offset) { |
| 6805 | const unsigned Opcode = IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END; |
| 6806 | const auto VTs = getVTList(MVT::Other); |
| 6807 | SDValue Ops[2] = { |
| 6808 | Chain, |
| 6809 | getFrameIndex(FrameIndex, |
| 6810 | getTargetLoweringInfo().getFrameIndexTy(getDataLayout()), |
| 6811 | true)}; |
| 6812 | |
| 6813 | FoldingSetNodeID ID; |
| 6814 | AddNodeIDNode(ID, Opcode, VTs, Ops); |
| 6815 | ID.AddInteger(FrameIndex); |
| 6816 | ID.AddInteger(Size); |
| 6817 | ID.AddInteger(Offset); |
| 6818 | void *IP = nullptr; |
| 6819 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) |
| 6820 | return SDValue(E, 0); |
| 6821 | |
| 6822 | LifetimeSDNode *N = newSDNode<LifetimeSDNode>( |
| 6823 | Opcode, dl.getIROrder(), dl.getDebugLoc(), VTs, Size, Offset); |
| 6824 | createOperands(N, Ops); |
| 6825 | CSEMap.InsertNode(N, IP); |
| 6826 | InsertNode(N); |
| 6827 | SDValue V(N, 0); |
| 6828 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 6829 | return V; |
| 6830 | } |
| 6831 | |
| 6832 | /// InferPointerInfo - If the specified ptr/offset is a frame index, infer a |
| 6833 | /// MachinePointerInfo record from it. This is particularly useful because the |
| 6834 | /// code generator has many cases where it doesn't bother passing in a |
| 6835 | /// MachinePointerInfo to getLoad or getStore when it has "FI+Cst". |
| 6836 | static MachinePointerInfo InferPointerInfo(const MachinePointerInfo &Info, |
| 6837 | SelectionDAG &DAG, SDValue Ptr, |
| 6838 | int64_t Offset = 0) { |
| 6839 | // If this is FI+Offset, we can model it. |
| 6840 | if (const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) |
| 6841 | return MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), |
| 6842 | FI->getIndex(), Offset); |
| 6843 | |
| 6844 | // If this is (FI+Offset1)+Offset2, we can model it. |
| 6845 | if (Ptr.getOpcode() != ISD::ADD || |
| 6846 | !isa<ConstantSDNode>(Ptr.getOperand(1)) || |
| 6847 | !isa<FrameIndexSDNode>(Ptr.getOperand(0))) |
| 6848 | return Info; |
| 6849 | |
| 6850 | int FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); |
| 6851 | return MachinePointerInfo::getFixedStack( |
| 6852 | DAG.getMachineFunction(), FI, |
| 6853 | Offset + cast<ConstantSDNode>(Ptr.getOperand(1))->getSExtValue()); |
| 6854 | } |
| 6855 | |
| 6856 | /// InferPointerInfo - If the specified ptr/offset is a frame index, infer a |
| 6857 | /// MachinePointerInfo record from it. This is particularly useful because the |
| 6858 | /// code generator has many cases where it doesn't bother passing in a |
| 6859 | /// MachinePointerInfo to getLoad or getStore when it has "FI+Cst". |
| 6860 | static MachinePointerInfo InferPointerInfo(const MachinePointerInfo &Info, |
| 6861 | SelectionDAG &DAG, SDValue Ptr, |
| 6862 | SDValue OffsetOp) { |
| 6863 | // If the 'Offset' value isn't a constant, we can't handle this. |
| 6864 | if (ConstantSDNode *OffsetNode = dyn_cast<ConstantSDNode>(OffsetOp)) |
| 6865 | return InferPointerInfo(Info, DAG, Ptr, OffsetNode->getSExtValue()); |
| 6866 | if (OffsetOp.isUndef()) |
| 6867 | return InferPointerInfo(Info, DAG, Ptr); |
| 6868 | return Info; |
| 6869 | } |
| 6870 | |
| 6871 | SDValue SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, |
| 6872 | EVT VT, const SDLoc &dl, SDValue Chain, |
| 6873 | SDValue Ptr, SDValue Offset, |
| 6874 | MachinePointerInfo PtrInfo, EVT MemVT, |
| 6875 | unsigned Alignment, |
| 6876 | MachineMemOperand::Flags MMOFlags, |
| 6877 | const AAMDNodes &AAInfo, const MDNode *Ranges) { |
| 6878 | assert(Chain.getValueType() == MVT::Other && |
| 6879 | "Invalid chain type" ); |
| 6880 | if (Alignment == 0) // Ensure that codegen never sees alignment 0 |
| 6881 | Alignment = getEVTAlignment(MemVT); |
| 6882 | |
| 6883 | MMOFlags |= MachineMemOperand::MOLoad; |
| 6884 | assert((MMOFlags & MachineMemOperand::MOStore) == 0); |
| 6885 | // If we don't have a PtrInfo, infer the trivial frame index case to simplify |
| 6886 | // clients. |
| 6887 | if (PtrInfo.V.isNull()) |
| 6888 | PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr, Offset); |
| 6889 | |
| 6890 | MachineFunction &MF = getMachineFunction(); |
| 6891 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
| 6892 | PtrInfo, MMOFlags, MemVT.getStoreSize(), Alignment, AAInfo, Ranges); |
| 6893 | return getLoad(AM, ExtType, VT, dl, Chain, Ptr, Offset, MemVT, MMO); |
| 6894 | } |
| 6895 | |
| 6896 | SDValue SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, |
| 6897 | EVT VT, const SDLoc &dl, SDValue Chain, |
| 6898 | SDValue Ptr, SDValue Offset, EVT MemVT, |
| 6899 | MachineMemOperand *MMO) { |
| 6900 | if (VT == MemVT) { |
| 6901 | ExtType = ISD::NON_EXTLOAD; |
| 6902 | } else if (ExtType == ISD::NON_EXTLOAD) { |
| 6903 | assert(VT == MemVT && "Non-extending load from different memory type!" ); |
| 6904 | } else { |
| 6905 | // Extending load. |
| 6906 | assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) && |
| 6907 | "Should only be an extending load, not truncating!" ); |
| 6908 | assert(VT.isInteger() == MemVT.isInteger() && |
| 6909 | "Cannot convert from FP to Int or Int -> FP!" ); |
| 6910 | assert(VT.isVector() == MemVT.isVector() && |
| 6911 | "Cannot use an ext load to convert to or from a vector!" ); |
| 6912 | assert((!VT.isVector() || |
| 6913 | VT.getVectorNumElements() == MemVT.getVectorNumElements()) && |
| 6914 | "Cannot use an ext load to change the number of vector elements!" ); |
| 6915 | } |
| 6916 | |
| 6917 | bool Indexed = AM != ISD::UNINDEXED; |
| 6918 | assert((Indexed || Offset.isUndef()) && "Unindexed load with an offset!" ); |
| 6919 | |
| 6920 | SDVTList VTs = Indexed ? |
| 6921 | getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other); |
| 6922 | SDValue Ops[] = { Chain, Ptr, Offset }; |
| 6923 | FoldingSetNodeID ID; |
| 6924 | AddNodeIDNode(ID, ISD::LOAD, VTs, Ops); |
| 6925 | ID.AddInteger(MemVT.getRawBits()); |
| 6926 | ID.AddInteger(getSyntheticNodeSubclassData<LoadSDNode>( |
| 6927 | dl.getIROrder(), VTs, AM, ExtType, MemVT, MMO)); |
| 6928 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 6929 | void *IP = nullptr; |
| 6930 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 6931 | cast<LoadSDNode>(E)->refineAlignment(MMO); |
| 6932 | return SDValue(E, 0); |
| 6933 | } |
| 6934 | auto *N = newSDNode<LoadSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, AM, |
| 6935 | ExtType, MemVT, MMO); |
| 6936 | createOperands(N, Ops); |
| 6937 | |
| 6938 | CSEMap.InsertNode(N, IP); |
| 6939 | InsertNode(N); |
| 6940 | SDValue V(N, 0); |
| 6941 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 6942 | return V; |
| 6943 | } |
| 6944 | |
| 6945 | SDValue SelectionDAG::getLoad(EVT VT, const SDLoc &dl, SDValue Chain, |
| 6946 | SDValue Ptr, MachinePointerInfo PtrInfo, |
| 6947 | unsigned Alignment, |
| 6948 | MachineMemOperand::Flags MMOFlags, |
| 6949 | const AAMDNodes &AAInfo, const MDNode *Ranges) { |
| 6950 | SDValue Undef = getUNDEF(Ptr.getValueType()); |
| 6951 | return getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Undef, |
| 6952 | PtrInfo, VT, Alignment, MMOFlags, AAInfo, Ranges); |
| 6953 | } |
| 6954 | |
| 6955 | SDValue SelectionDAG::getLoad(EVT VT, const SDLoc &dl, SDValue Chain, |
| 6956 | SDValue Ptr, MachineMemOperand *MMO) { |
| 6957 | SDValue Undef = getUNDEF(Ptr.getValueType()); |
| 6958 | return getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Undef, |
| 6959 | VT, MMO); |
| 6960 | } |
| 6961 | |
| 6962 | SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, const SDLoc &dl, |
| 6963 | EVT VT, SDValue Chain, SDValue Ptr, |
| 6964 | MachinePointerInfo PtrInfo, EVT MemVT, |
| 6965 | unsigned Alignment, |
| 6966 | MachineMemOperand::Flags MMOFlags, |
| 6967 | const AAMDNodes &AAInfo) { |
| 6968 | SDValue Undef = getUNDEF(Ptr.getValueType()); |
| 6969 | return getLoad(ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Undef, PtrInfo, |
| 6970 | MemVT, Alignment, MMOFlags, AAInfo); |
| 6971 | } |
| 6972 | |
| 6973 | SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, const SDLoc &dl, |
| 6974 | EVT VT, SDValue Chain, SDValue Ptr, EVT MemVT, |
| 6975 | MachineMemOperand *MMO) { |
| 6976 | SDValue Undef = getUNDEF(Ptr.getValueType()); |
| 6977 | return getLoad(ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Undef, |
| 6978 | MemVT, MMO); |
| 6979 | } |
| 6980 | |
| 6981 | SDValue SelectionDAG::getIndexedLoad(SDValue OrigLoad, const SDLoc &dl, |
| 6982 | SDValue Base, SDValue Offset, |
| 6983 | ISD::MemIndexedMode AM) { |
| 6984 | LoadSDNode *LD = cast<LoadSDNode>(OrigLoad); |
| 6985 | assert(LD->getOffset().isUndef() && "Load is already a indexed load!" ); |
| 6986 | // Don't propagate the invariant or dereferenceable flags. |
| 6987 | auto MMOFlags = |
| 6988 | LD->getMemOperand()->getFlags() & |
| 6989 | ~(MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable); |
| 6990 | return getLoad(AM, LD->getExtensionType(), OrigLoad.getValueType(), dl, |
| 6991 | LD->getChain(), Base, Offset, LD->getPointerInfo(), |
| 6992 | LD->getMemoryVT(), LD->getAlignment(), MMOFlags, |
| 6993 | LD->getAAInfo()); |
| 6994 | } |
| 6995 | |
| 6996 | SDValue SelectionDAG::getStore(SDValue Chain, const SDLoc &dl, SDValue Val, |
| 6997 | SDValue Ptr, MachinePointerInfo PtrInfo, |
| 6998 | unsigned Alignment, |
| 6999 | MachineMemOperand::Flags MMOFlags, |
| 7000 | const AAMDNodes &AAInfo) { |
| 7001 | assert(Chain.getValueType() == MVT::Other && "Invalid chain type" ); |
| 7002 | if (Alignment == 0) // Ensure that codegen never sees alignment 0 |
| 7003 | Alignment = getEVTAlignment(Val.getValueType()); |
| 7004 | |
| 7005 | MMOFlags |= MachineMemOperand::MOStore; |
| 7006 | assert((MMOFlags & MachineMemOperand::MOLoad) == 0); |
| 7007 | |
| 7008 | if (PtrInfo.V.isNull()) |
| 7009 | PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr); |
| 7010 | |
| 7011 | MachineFunction &MF = getMachineFunction(); |
| 7012 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
| 7013 | PtrInfo, MMOFlags, Val.getValueType().getStoreSize(), Alignment, AAInfo); |
| 7014 | return getStore(Chain, dl, Val, Ptr, MMO); |
| 7015 | } |
| 7016 | |
| 7017 | SDValue SelectionDAG::getStore(SDValue Chain, const SDLoc &dl, SDValue Val, |
| 7018 | SDValue Ptr, MachineMemOperand *MMO) { |
| 7019 | assert(Chain.getValueType() == MVT::Other && |
| 7020 | "Invalid chain type" ); |
| 7021 | EVT VT = Val.getValueType(); |
| 7022 | SDVTList VTs = getVTList(MVT::Other); |
| 7023 | SDValue Undef = getUNDEF(Ptr.getValueType()); |
| 7024 | SDValue Ops[] = { Chain, Val, Ptr, Undef }; |
| 7025 | FoldingSetNodeID ID; |
| 7026 | AddNodeIDNode(ID, ISD::STORE, VTs, Ops); |
| 7027 | ID.AddInteger(VT.getRawBits()); |
| 7028 | ID.AddInteger(getSyntheticNodeSubclassData<StoreSDNode>( |
| 7029 | dl.getIROrder(), VTs, ISD::UNINDEXED, false, VT, MMO)); |
| 7030 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 7031 | void *IP = nullptr; |
| 7032 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 7033 | cast<StoreSDNode>(E)->refineAlignment(MMO); |
| 7034 | return SDValue(E, 0); |
| 7035 | } |
| 7036 | auto *N = newSDNode<StoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, |
| 7037 | ISD::UNINDEXED, false, VT, MMO); |
| 7038 | createOperands(N, Ops); |
| 7039 | |
| 7040 | CSEMap.InsertNode(N, IP); |
| 7041 | InsertNode(N); |
| 7042 | SDValue V(N, 0); |
| 7043 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7044 | return V; |
| 7045 | } |
| 7046 | |
| 7047 | SDValue SelectionDAG::getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val, |
| 7048 | SDValue Ptr, MachinePointerInfo PtrInfo, |
| 7049 | EVT SVT, unsigned Alignment, |
| 7050 | MachineMemOperand::Flags MMOFlags, |
| 7051 | const AAMDNodes &AAInfo) { |
| 7052 | assert(Chain.getValueType() == MVT::Other && |
| 7053 | "Invalid chain type" ); |
| 7054 | if (Alignment == 0) // Ensure that codegen never sees alignment 0 |
| 7055 | Alignment = getEVTAlignment(SVT); |
| 7056 | |
| 7057 | MMOFlags |= MachineMemOperand::MOStore; |
| 7058 | assert((MMOFlags & MachineMemOperand::MOLoad) == 0); |
| 7059 | |
| 7060 | if (PtrInfo.V.isNull()) |
| 7061 | PtrInfo = InferPointerInfo(PtrInfo, *this, Ptr); |
| 7062 | |
| 7063 | MachineFunction &MF = getMachineFunction(); |
| 7064 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
| 7065 | PtrInfo, MMOFlags, SVT.getStoreSize(), Alignment, AAInfo); |
| 7066 | return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO); |
| 7067 | } |
| 7068 | |
| 7069 | SDValue SelectionDAG::getTruncStore(SDValue Chain, const SDLoc &dl, SDValue Val, |
| 7070 | SDValue Ptr, EVT SVT, |
| 7071 | MachineMemOperand *MMO) { |
| 7072 | EVT VT = Val.getValueType(); |
| 7073 | |
| 7074 | assert(Chain.getValueType() == MVT::Other && |
| 7075 | "Invalid chain type" ); |
| 7076 | if (VT == SVT) |
| 7077 | return getStore(Chain, dl, Val, Ptr, MMO); |
| 7078 | |
| 7079 | assert(SVT.getScalarType().bitsLT(VT.getScalarType()) && |
| 7080 | "Should only be a truncating store, not extending!" ); |
| 7081 | assert(VT.isInteger() == SVT.isInteger() && |
| 7082 | "Can't do FP-INT conversion!" ); |
| 7083 | assert(VT.isVector() == SVT.isVector() && |
| 7084 | "Cannot use trunc store to convert to or from a vector!" ); |
| 7085 | assert((!VT.isVector() || |
| 7086 | VT.getVectorNumElements() == SVT.getVectorNumElements()) && |
| 7087 | "Cannot use trunc store to change the number of vector elements!" ); |
| 7088 | |
| 7089 | SDVTList VTs = getVTList(MVT::Other); |
| 7090 | SDValue Undef = getUNDEF(Ptr.getValueType()); |
| 7091 | SDValue Ops[] = { Chain, Val, Ptr, Undef }; |
| 7092 | FoldingSetNodeID ID; |
| 7093 | AddNodeIDNode(ID, ISD::STORE, VTs, Ops); |
| 7094 | ID.AddInteger(SVT.getRawBits()); |
| 7095 | ID.AddInteger(getSyntheticNodeSubclassData<StoreSDNode>( |
| 7096 | dl.getIROrder(), VTs, ISD::UNINDEXED, true, SVT, MMO)); |
| 7097 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 7098 | void *IP = nullptr; |
| 7099 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 7100 | cast<StoreSDNode>(E)->refineAlignment(MMO); |
| 7101 | return SDValue(E, 0); |
| 7102 | } |
| 7103 | auto *N = newSDNode<StoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, |
| 7104 | ISD::UNINDEXED, true, SVT, MMO); |
| 7105 | createOperands(N, Ops); |
| 7106 | |
| 7107 | CSEMap.InsertNode(N, IP); |
| 7108 | InsertNode(N); |
| 7109 | SDValue V(N, 0); |
| 7110 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7111 | return V; |
| 7112 | } |
| 7113 | |
| 7114 | SDValue SelectionDAG::getIndexedStore(SDValue OrigStore, const SDLoc &dl, |
| 7115 | SDValue Base, SDValue Offset, |
| 7116 | ISD::MemIndexedMode AM) { |
| 7117 | StoreSDNode *ST = cast<StoreSDNode>(OrigStore); |
| 7118 | assert(ST->getOffset().isUndef() && "Store is already a indexed store!" ); |
| 7119 | SDVTList VTs = getVTList(Base.getValueType(), MVT::Other); |
| 7120 | SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset }; |
| 7121 | FoldingSetNodeID ID; |
| 7122 | AddNodeIDNode(ID, ISD::STORE, VTs, Ops); |
| 7123 | ID.AddInteger(ST->getMemoryVT().getRawBits()); |
| 7124 | ID.AddInteger(ST->getRawSubclassData()); |
| 7125 | ID.AddInteger(ST->getPointerInfo().getAddrSpace()); |
| 7126 | void *IP = nullptr; |
| 7127 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) |
| 7128 | return SDValue(E, 0); |
| 7129 | |
| 7130 | auto *N = newSDNode<StoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, AM, |
| 7131 | ST->isTruncatingStore(), ST->getMemoryVT(), |
| 7132 | ST->getMemOperand()); |
| 7133 | createOperands(N, Ops); |
| 7134 | |
| 7135 | CSEMap.InsertNode(N, IP); |
| 7136 | InsertNode(N); |
| 7137 | SDValue V(N, 0); |
| 7138 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7139 | return V; |
| 7140 | } |
| 7141 | |
| 7142 | SDValue SelectionDAG::getMaskedLoad(EVT VT, const SDLoc &dl, SDValue Chain, |
| 7143 | SDValue Ptr, SDValue Mask, SDValue PassThru, |
| 7144 | EVT MemVT, MachineMemOperand *MMO, |
| 7145 | ISD::LoadExtType ExtTy, bool isExpanding) { |
| 7146 | SDVTList VTs = getVTList(VT, MVT::Other); |
| 7147 | SDValue Ops[] = { Chain, Ptr, Mask, PassThru }; |
| 7148 | FoldingSetNodeID ID; |
| 7149 | AddNodeIDNode(ID, ISD::MLOAD, VTs, Ops); |
| 7150 | ID.AddInteger(VT.getRawBits()); |
| 7151 | ID.AddInteger(getSyntheticNodeSubclassData<MaskedLoadSDNode>( |
| 7152 | dl.getIROrder(), VTs, ExtTy, isExpanding, MemVT, MMO)); |
| 7153 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 7154 | void *IP = nullptr; |
| 7155 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 7156 | cast<MaskedLoadSDNode>(E)->refineAlignment(MMO); |
| 7157 | return SDValue(E, 0); |
| 7158 | } |
| 7159 | auto *N = newSDNode<MaskedLoadSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, |
| 7160 | ExtTy, isExpanding, MemVT, MMO); |
| 7161 | createOperands(N, Ops); |
| 7162 | |
| 7163 | CSEMap.InsertNode(N, IP); |
| 7164 | InsertNode(N); |
| 7165 | SDValue V(N, 0); |
| 7166 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7167 | return V; |
| 7168 | } |
| 7169 | |
| 7170 | SDValue SelectionDAG::getMaskedStore(SDValue Chain, const SDLoc &dl, |
| 7171 | SDValue Val, SDValue Ptr, SDValue Mask, |
| 7172 | EVT MemVT, MachineMemOperand *MMO, |
| 7173 | bool IsTruncating, bool IsCompressing) { |
| 7174 | assert(Chain.getValueType() == MVT::Other && |
| 7175 | "Invalid chain type" ); |
| 7176 | EVT VT = Val.getValueType(); |
| 7177 | SDVTList VTs = getVTList(MVT::Other); |
| 7178 | SDValue Ops[] = { Chain, Val, Ptr, Mask }; |
| 7179 | FoldingSetNodeID ID; |
| 7180 | AddNodeIDNode(ID, ISD::MSTORE, VTs, Ops); |
| 7181 | ID.AddInteger(VT.getRawBits()); |
| 7182 | ID.AddInteger(getSyntheticNodeSubclassData<MaskedStoreSDNode>( |
| 7183 | dl.getIROrder(), VTs, IsTruncating, IsCompressing, MemVT, MMO)); |
| 7184 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 7185 | void *IP = nullptr; |
| 7186 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 7187 | cast<MaskedStoreSDNode>(E)->refineAlignment(MMO); |
| 7188 | return SDValue(E, 0); |
| 7189 | } |
| 7190 | auto *N = newSDNode<MaskedStoreSDNode>(dl.getIROrder(), dl.getDebugLoc(), VTs, |
| 7191 | IsTruncating, IsCompressing, MemVT, MMO); |
| 7192 | createOperands(N, Ops); |
| 7193 | |
| 7194 | CSEMap.InsertNode(N, IP); |
| 7195 | InsertNode(N); |
| 7196 | SDValue V(N, 0); |
| 7197 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7198 | return V; |
| 7199 | } |
| 7200 | |
| 7201 | SDValue SelectionDAG::getMaskedGather(SDVTList VTs, EVT VT, const SDLoc &dl, |
| 7202 | ArrayRef<SDValue> Ops, |
| 7203 | MachineMemOperand *MMO) { |
| 7204 | assert(Ops.size() == 6 && "Incompatible number of operands" ); |
| 7205 | |
| 7206 | FoldingSetNodeID ID; |
| 7207 | AddNodeIDNode(ID, ISD::MGATHER, VTs, Ops); |
| 7208 | ID.AddInteger(VT.getRawBits()); |
| 7209 | ID.AddInteger(getSyntheticNodeSubclassData<MaskedGatherSDNode>( |
| 7210 | dl.getIROrder(), VTs, VT, MMO)); |
| 7211 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 7212 | void *IP = nullptr; |
| 7213 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 7214 | cast<MaskedGatherSDNode>(E)->refineAlignment(MMO); |
| 7215 | return SDValue(E, 0); |
| 7216 | } |
| 7217 | |
| 7218 | auto *N = newSDNode<MaskedGatherSDNode>(dl.getIROrder(), dl.getDebugLoc(), |
| 7219 | VTs, VT, MMO); |
| 7220 | createOperands(N, Ops); |
| 7221 | |
| 7222 | assert(N->getPassThru().getValueType() == N->getValueType(0) && |
| 7223 | "Incompatible type of the PassThru value in MaskedGatherSDNode" ); |
| 7224 | assert(N->getMask().getValueType().getVectorNumElements() == |
| 7225 | N->getValueType(0).getVectorNumElements() && |
| 7226 | "Vector width mismatch between mask and data" ); |
| 7227 | assert(N->getIndex().getValueType().getVectorNumElements() >= |
| 7228 | N->getValueType(0).getVectorNumElements() && |
| 7229 | "Vector width mismatch between index and data" ); |
| 7230 | assert(isa<ConstantSDNode>(N->getScale()) && |
| 7231 | cast<ConstantSDNode>(N->getScale())->getAPIntValue().isPowerOf2() && |
| 7232 | "Scale should be a constant power of 2" ); |
| 7233 | |
| 7234 | CSEMap.InsertNode(N, IP); |
| 7235 | InsertNode(N); |
| 7236 | SDValue V(N, 0); |
| 7237 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7238 | return V; |
| 7239 | } |
| 7240 | |
| 7241 | SDValue SelectionDAG::getMaskedScatter(SDVTList VTs, EVT VT, const SDLoc &dl, |
| 7242 | ArrayRef<SDValue> Ops, |
| 7243 | MachineMemOperand *MMO) { |
| 7244 | assert(Ops.size() == 6 && "Incompatible number of operands" ); |
| 7245 | |
| 7246 | FoldingSetNodeID ID; |
| 7247 | AddNodeIDNode(ID, ISD::MSCATTER, VTs, Ops); |
| 7248 | ID.AddInteger(VT.getRawBits()); |
| 7249 | ID.AddInteger(getSyntheticNodeSubclassData<MaskedScatterSDNode>( |
| 7250 | dl.getIROrder(), VTs, VT, MMO)); |
| 7251 | ID.AddInteger(MMO->getPointerInfo().getAddrSpace()); |
| 7252 | void *IP = nullptr; |
| 7253 | if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) { |
| 7254 | cast<MaskedScatterSDNode>(E)->refineAlignment(MMO); |
| 7255 | return SDValue(E, 0); |
| 7256 | } |
| 7257 | auto *N = newSDNode<MaskedScatterSDNode>(dl.getIROrder(), dl.getDebugLoc(), |
| 7258 | VTs, VT, MMO); |
| 7259 | createOperands(N, Ops); |
| 7260 | |
| 7261 | assert(N->getMask().getValueType().getVectorNumElements() == |
| 7262 | N->getValue().getValueType().getVectorNumElements() && |
| 7263 | "Vector width mismatch between mask and data" ); |
| 7264 | assert(N->getIndex().getValueType().getVectorNumElements() >= |
| 7265 | N->getValue().getValueType().getVectorNumElements() && |
| 7266 | "Vector width mismatch between index and data" ); |
| 7267 | assert(isa<ConstantSDNode>(N->getScale()) && |
| 7268 | cast<ConstantSDNode>(N->getScale())->getAPIntValue().isPowerOf2() && |
| 7269 | "Scale should be a constant power of 2" ); |
| 7270 | |
| 7271 | CSEMap.InsertNode(N, IP); |
| 7272 | InsertNode(N); |
| 7273 | SDValue V(N, 0); |
| 7274 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7275 | return V; |
| 7276 | } |
| 7277 | |
| 7278 | SDValue SelectionDAG::simplifySelect(SDValue Cond, SDValue T, SDValue F) { |
| 7279 | // select undef, T, F --> T (if T is a constant), otherwise F |
| 7280 | // select, ?, undef, F --> F |
| 7281 | // select, ?, T, undef --> T |
| 7282 | if (Cond.isUndef()) |
| 7283 | return isConstantValueOfAnyType(T) ? T : F; |
| 7284 | if (T.isUndef()) |
| 7285 | return F; |
| 7286 | if (F.isUndef()) |
| 7287 | return T; |
| 7288 | |
| 7289 | // select true, T, F --> T |
| 7290 | // select false, T, F --> F |
| 7291 | if (auto *CondC = dyn_cast<ConstantSDNode>(Cond)) |
| 7292 | return CondC->isNullValue() ? F : T; |
| 7293 | |
| 7294 | // TODO: This should simplify VSELECT with constant condition using something |
| 7295 | // like this (but check boolean contents to be complete?): |
| 7296 | // if (ISD::isBuildVectorAllOnes(Cond.getNode())) |
| 7297 | // return T; |
| 7298 | // if (ISD::isBuildVectorAllZeros(Cond.getNode())) |
| 7299 | // return F; |
| 7300 | |
| 7301 | // select ?, T, T --> T |
| 7302 | if (T == F) |
| 7303 | return T; |
| 7304 | |
| 7305 | return SDValue(); |
| 7306 | } |
| 7307 | |
| 7308 | SDValue SelectionDAG::simplifyShift(SDValue X, SDValue Y) { |
| 7309 | // shift undef, Y --> 0 (can always assume that the undef value is 0) |
| 7310 | if (X.isUndef()) |
| 7311 | return getConstant(0, SDLoc(X.getNode()), X.getValueType()); |
| 7312 | // shift X, undef --> undef (because it may shift by the bitwidth) |
| 7313 | if (Y.isUndef()) |
| 7314 | return getUNDEF(X.getValueType()); |
| 7315 | |
| 7316 | // shift 0, Y --> 0 |
| 7317 | // shift X, 0 --> X |
| 7318 | if (isNullOrNullSplat(X) || isNullOrNullSplat(Y)) |
| 7319 | return X; |
| 7320 | |
| 7321 | // shift X, C >= bitwidth(X) --> undef |
| 7322 | // All vector elements must be too big (or undef) to avoid partial undefs. |
| 7323 | auto isShiftTooBig = [X](ConstantSDNode *Val) { |
| 7324 | return !Val || Val->getAPIntValue().uge(X.getScalarValueSizeInBits()); |
| 7325 | }; |
| 7326 | if (ISD::matchUnaryPredicate(Y, isShiftTooBig, true)) |
| 7327 | return getUNDEF(X.getValueType()); |
| 7328 | |
| 7329 | return SDValue(); |
| 7330 | } |
| 7331 | |
| 7332 | // TODO: Use fast-math-flags to enable more simplifications. |
| 7333 | SDValue SelectionDAG::simplifyFPBinop(unsigned Opcode, SDValue X, SDValue Y) { |
| 7334 | ConstantFPSDNode *YC = isConstOrConstSplatFP(Y, /* AllowUndefs */ true); |
| 7335 | if (!YC) |
| 7336 | return SDValue(); |
| 7337 | |
| 7338 | // X + -0.0 --> X |
| 7339 | if (Opcode == ISD::FADD) |
| 7340 | if (YC->getValueAPF().isNegZero()) |
| 7341 | return X; |
| 7342 | |
| 7343 | // X - +0.0 --> X |
| 7344 | if (Opcode == ISD::FSUB) |
| 7345 | if (YC->getValueAPF().isPosZero()) |
| 7346 | return X; |
| 7347 | |
| 7348 | // X * 1.0 --> X |
| 7349 | // X / 1.0 --> X |
| 7350 | if (Opcode == ISD::FMUL || Opcode == ISD::FDIV) |
| 7351 | if (YC->getValueAPF().isExactlyValue(1.0)) |
| 7352 | return X; |
| 7353 | |
| 7354 | return SDValue(); |
| 7355 | } |
| 7356 | |
| 7357 | SDValue SelectionDAG::getVAArg(EVT VT, const SDLoc &dl, SDValue Chain, |
| 7358 | SDValue Ptr, SDValue SV, unsigned Align) { |
| 7359 | SDValue Ops[] = { Chain, Ptr, SV, getTargetConstant(Align, dl, MVT::i32) }; |
| 7360 | return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops); |
| 7361 | } |
| 7362 | |
| 7363 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, |
| 7364 | ArrayRef<SDUse> Ops) { |
| 7365 | switch (Ops.size()) { |
| 7366 | case 0: return getNode(Opcode, DL, VT); |
| 7367 | case 1: return getNode(Opcode, DL, VT, static_cast<const SDValue>(Ops[0])); |
| 7368 | case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]); |
| 7369 | case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]); |
| 7370 | default: break; |
| 7371 | } |
| 7372 | |
| 7373 | // Copy from an SDUse array into an SDValue array for use with |
| 7374 | // the regular getNode logic. |
| 7375 | SmallVector<SDValue, 8> NewOps(Ops.begin(), Ops.end()); |
| 7376 | return getNode(Opcode, DL, VT, NewOps); |
| 7377 | } |
| 7378 | |
| 7379 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, |
| 7380 | ArrayRef<SDValue> Ops, const SDNodeFlags Flags) { |
| 7381 | unsigned NumOps = Ops.size(); |
| 7382 | switch (NumOps) { |
| 7383 | case 0: return getNode(Opcode, DL, VT); |
| 7384 | case 1: return getNode(Opcode, DL, VT, Ops[0], Flags); |
| 7385 | case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Flags); |
| 7386 | case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2], Flags); |
| 7387 | default: break; |
| 7388 | } |
| 7389 | |
| 7390 | switch (Opcode) { |
| 7391 | default: break; |
| 7392 | case ISD::BUILD_VECTOR: |
| 7393 | // Attempt to simplify BUILD_VECTOR. |
| 7394 | if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this)) |
| 7395 | return V; |
| 7396 | break; |
| 7397 | case ISD::CONCAT_VECTORS: |
| 7398 | if (SDValue V = foldCONCAT_VECTORS(DL, VT, Ops, *this)) |
| 7399 | return V; |
| 7400 | break; |
| 7401 | case ISD::SELECT_CC: |
| 7402 | assert(NumOps == 5 && "SELECT_CC takes 5 operands!" ); |
| 7403 | assert(Ops[0].getValueType() == Ops[1].getValueType() && |
| 7404 | "LHS and RHS of condition must have same type!" ); |
| 7405 | assert(Ops[2].getValueType() == Ops[3].getValueType() && |
| 7406 | "True and False arms of SelectCC must have same type!" ); |
| 7407 | assert(Ops[2].getValueType() == VT && |
| 7408 | "select_cc node must be of same type as true and false value!" ); |
| 7409 | break; |
| 7410 | case ISD::BR_CC: |
| 7411 | assert(NumOps == 5 && "BR_CC takes 5 operands!" ); |
| 7412 | assert(Ops[2].getValueType() == Ops[3].getValueType() && |
| 7413 | "LHS/RHS of comparison should match types!" ); |
| 7414 | break; |
| 7415 | } |
| 7416 | |
| 7417 | // Memoize nodes. |
| 7418 | SDNode *N; |
| 7419 | SDVTList VTs = getVTList(VT); |
| 7420 | |
| 7421 | if (VT != MVT::Glue) { |
| 7422 | FoldingSetNodeID ID; |
| 7423 | AddNodeIDNode(ID, Opcode, VTs, Ops); |
| 7424 | void *IP = nullptr; |
| 7425 | |
| 7426 | if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) |
| 7427 | return SDValue(E, 0); |
| 7428 | |
| 7429 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 7430 | createOperands(N, Ops); |
| 7431 | |
| 7432 | CSEMap.InsertNode(N, IP); |
| 7433 | } else { |
| 7434 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 7435 | createOperands(N, Ops); |
| 7436 | } |
| 7437 | |
| 7438 | InsertNode(N); |
| 7439 | SDValue V(N, 0); |
| 7440 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7441 | return V; |
| 7442 | } |
| 7443 | |
| 7444 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, |
| 7445 | ArrayRef<EVT> ResultTys, ArrayRef<SDValue> Ops) { |
| 7446 | return getNode(Opcode, DL, getVTList(ResultTys), Ops); |
| 7447 | } |
| 7448 | |
| 7449 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, |
| 7450 | ArrayRef<SDValue> Ops) { |
| 7451 | if (VTList.NumVTs == 1) |
| 7452 | return getNode(Opcode, DL, VTList.VTs[0], Ops); |
| 7453 | |
| 7454 | #if 0 |
| 7455 | switch (Opcode) { |
| 7456 | // FIXME: figure out how to safely handle things like |
| 7457 | // int foo(int x) { return 1 << (x & 255); } |
| 7458 | // int bar() { return foo(256); } |
| 7459 | case ISD::SRA_PARTS: |
| 7460 | case ISD::SRL_PARTS: |
| 7461 | case ISD::SHL_PARTS: |
| 7462 | if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG && |
| 7463 | cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1) |
| 7464 | return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0)); |
| 7465 | else if (N3.getOpcode() == ISD::AND) |
| 7466 | if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) { |
| 7467 | // If the and is only masking out bits that cannot effect the shift, |
| 7468 | // eliminate the and. |
| 7469 | unsigned NumBits = VT.getScalarSizeInBits()*2; |
| 7470 | if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1) |
| 7471 | return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0)); |
| 7472 | } |
| 7473 | break; |
| 7474 | } |
| 7475 | #endif |
| 7476 | |
| 7477 | // Memoize the node unless it returns a flag. |
| 7478 | SDNode *N; |
| 7479 | if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) { |
| 7480 | FoldingSetNodeID ID; |
| 7481 | AddNodeIDNode(ID, Opcode, VTList, Ops); |
| 7482 | void *IP = nullptr; |
| 7483 | if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) |
| 7484 | return SDValue(E, 0); |
| 7485 | |
| 7486 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTList); |
| 7487 | createOperands(N, Ops); |
| 7488 | CSEMap.InsertNode(N, IP); |
| 7489 | } else { |
| 7490 | N = newSDNode<SDNode>(Opcode, DL.getIROrder(), DL.getDebugLoc(), VTList); |
| 7491 | createOperands(N, Ops); |
| 7492 | } |
| 7493 | InsertNode(N); |
| 7494 | SDValue V(N, 0); |
| 7495 | NewSDValueDbgMsg(V, "Creating new node: " , this); |
| 7496 | return V; |
| 7497 | } |
| 7498 | |
| 7499 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, |
| 7500 | SDVTList VTList) { |
| 7501 | return getNode(Opcode, DL, VTList, None); |
| 7502 | } |
| 7503 | |
| 7504 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, |
| 7505 | SDValue N1) { |
| 7506 | SDValue Ops[] = { N1 }; |
| 7507 | return getNode(Opcode, DL, VTList, Ops); |
| 7508 | } |
| 7509 | |
| 7510 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, |
| 7511 | SDValue N1, SDValue N2) { |
| 7512 | SDValue Ops[] = { N1, N2 }; |
| 7513 | return getNode(Opcode, DL, VTList, Ops); |
| 7514 | } |
| 7515 | |
| 7516 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, |
| 7517 | SDValue N1, SDValue N2, SDValue N3) { |
| 7518 | SDValue Ops[] = { N1, N2, N3 }; |
| 7519 | return getNode(Opcode, DL, VTList, Ops); |
| 7520 | } |
| 7521 | |
| 7522 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, |
| 7523 | SDValue N1, SDValue N2, SDValue N3, SDValue N4) { |
| 7524 | SDValue Ops[] = { N1, N2, N3, N4 }; |
| 7525 | return getNode(Opcode, DL, VTList, Ops); |
| 7526 | } |
| 7527 | |
| 7528 | SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, |
| 7529 | SDValue N1, SDValue N2, SDValue N3, SDValue N4, |
| 7530 | SDValue N5) { |
| 7531 | SDValue Ops[] = { N1, N2, N3, N4, N5 }; |
| 7532 | return getNode(Opcode, DL, VTList, Ops); |
| 7533 | } |
| 7534 | |
| 7535 | SDVTList SelectionDAG::getVTList(EVT VT) { |
| 7536 | return makeVTList(SDNode::getValueTypeList(VT), 1); |
| 7537 | } |
| 7538 | |
| 7539 | SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) { |
| 7540 | FoldingSetNodeID ID; |
| 7541 | ID.AddInteger(2U); |
| 7542 | ID.AddInteger(VT1.getRawBits()); |
| 7543 | ID.AddInteger(VT2.getRawBits()); |
| 7544 | |
| 7545 | void *IP = nullptr; |
| 7546 | SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, IP); |
| 7547 | if (!Result) { |
| 7548 | EVT *Array = Allocator.Allocate<EVT>(2); |
| 7549 | Array[0] = VT1; |
| 7550 | Array[1] = VT2; |
| 7551 | Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 2); |
| 7552 | VTListMap.InsertNode(Result, IP); |
| 7553 | } |
| 7554 | return Result->getSDVTList(); |
| 7555 | } |
| 7556 | |
| 7557 | SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) { |
| 7558 | FoldingSetNodeID ID; |
| 7559 | ID.AddInteger(3U); |
| 7560 | ID.AddInteger(VT1.getRawBits()); |
| 7561 | ID.AddInteger(VT2.getRawBits()); |
| 7562 | ID.AddInteger(VT3.getRawBits()); |
| 7563 | |
| 7564 | void *IP = nullptr; |
| 7565 | SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, IP); |
| 7566 | if (!Result) { |
| 7567 | EVT *Array = Allocator.Allocate<EVT>(3); |
| 7568 | Array[0] = VT1; |
| 7569 | Array[1] = VT2; |
| 7570 | Array[2] = VT3; |
| 7571 | Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 3); |
| 7572 | VTListMap.InsertNode(Result, IP); |
| 7573 | } |
| 7574 | return Result->getSDVTList(); |
| 7575 | } |
| 7576 | |
| 7577 | SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) { |
| 7578 | FoldingSetNodeID ID; |
| 7579 | ID.AddInteger(4U); |
| 7580 | ID.AddInteger(VT1.getRawBits()); |
| 7581 | ID.AddInteger(VT2.getRawBits()); |
| 7582 | ID.AddInteger(VT3.getRawBits()); |
| 7583 | ID.AddInteger(VT4.getRawBits()); |
| 7584 | |
| 7585 | void *IP = nullptr; |
| 7586 | SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, IP); |
| 7587 | if (!Result) { |
| 7588 | EVT *Array = Allocator.Allocate<EVT>(4); |
| 7589 | Array[0] = VT1; |
| 7590 | Array[1] = VT2; |
| 7591 | Array[2] = VT3; |
| 7592 | Array[3] = VT4; |
| 7593 | Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, 4); |
| 7594 | VTListMap.InsertNode(Result, IP); |
| 7595 | } |
| 7596 | return Result->getSDVTList(); |
| 7597 | } |
| 7598 | |
| 7599 | SDVTList SelectionDAG::getVTList(ArrayRef<EVT> VTs) { |
| 7600 | unsigned NumVTs = VTs.size(); |
| 7601 | FoldingSetNodeID ID; |
| 7602 | ID.AddInteger(NumVTs); |
| 7603 | for (unsigned index = 0; index < NumVTs; index++) { |
| 7604 | ID.AddInteger(VTs[index].getRawBits()); |
| 7605 | } |
| 7606 | |
| 7607 | void *IP = nullptr; |
| 7608 | SDVTListNode *Result = VTListMap.FindNodeOrInsertPos(ID, IP); |
| 7609 | if (!Result) { |
| 7610 | EVT *Array = Allocator.Allocate<EVT>(NumVTs); |
| 7611 | llvm::copy(VTs, Array); |
| 7612 | Result = new (Allocator) SDVTListNode(ID.Intern(Allocator), Array, NumVTs); |
| 7613 | VTListMap.InsertNode(Result, IP); |
| 7614 | } |
| 7615 | return Result->getSDVTList(); |
| 7616 | } |
| 7617 | |
| 7618 | |
| 7619 | /// UpdateNodeOperands - *Mutate* the specified node in-place to have the |
| 7620 | /// specified operands. If the resultant node already exists in the DAG, |
| 7621 | /// this does not modify the specified node, instead it returns the node that |
| 7622 | /// already exists. If the resultant node does not exist in the DAG, the |
| 7623 | /// input node is returned. As a degenerate case, if you specify the same |
| 7624 | /// input operands as the node already has, the input node is returned. |
| 7625 | SDNode *SelectionDAG::UpdateNodeOperands(SDNode *N, SDValue Op) { |
| 7626 | assert(N->getNumOperands() == 1 && "Update with wrong number of operands" ); |
| 7627 | |
| 7628 | // Check to see if there is no change. |
| 7629 | if (Op == N->getOperand(0)) return N; |
| 7630 | |
| 7631 | // See if the modified node already exists. |
| 7632 | void *InsertPos = nullptr; |
| 7633 | if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos)) |
| 7634 | return Existing; |
| 7635 | |
| 7636 | // Nope it doesn't. Remove the node from its current place in the maps. |
| 7637 | if (InsertPos) |
| 7638 | if (!RemoveNodeFromCSEMaps(N)) |
| 7639 | InsertPos = nullptr; |
| 7640 | |
| 7641 | // Now we update the operands. |
| 7642 | N->OperandList[0].set(Op); |
| 7643 | |
| 7644 | updateDivergence(N); |
| 7645 | // If this gets put into a CSE map, add it. |
| 7646 | if (InsertPos) CSEMap.InsertNode(N, InsertPos); |
| 7647 | return N; |
| 7648 | } |
| 7649 | |
| 7650 | SDNode *SelectionDAG::UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2) { |
| 7651 | assert(N->getNumOperands() == 2 && "Update with wrong number of operands" ); |
| 7652 | |
| 7653 | // Check to see if there is no change. |
| 7654 | if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1)) |
| 7655 | return N; // No operands changed, just return the input node. |
| 7656 | |
| 7657 | // See if the modified node already exists. |
| 7658 | void *InsertPos = nullptr; |
| 7659 | if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos)) |
| 7660 | return Existing; |
| 7661 | |
| 7662 | // Nope it doesn't. Remove the node from its current place in the maps. |
| 7663 | if (InsertPos) |
| 7664 | if (!RemoveNodeFromCSEMaps(N)) |
| 7665 | InsertPos = nullptr; |
| 7666 | |
| 7667 | // Now we update the operands. |
| 7668 | if (N->OperandList[0] != Op1) |
| 7669 | N->OperandList[0].set(Op1); |
| 7670 | if (N->OperandList[1] != Op2) |
| 7671 | N->OperandList[1].set(Op2); |
| 7672 | |
| 7673 | updateDivergence(N); |
| 7674 | // If this gets put into a CSE map, add it. |
| 7675 | if (InsertPos) CSEMap.InsertNode(N, InsertPos); |
| 7676 | return N; |
| 7677 | } |
| 7678 | |
| 7679 | SDNode *SelectionDAG:: |
| 7680 | UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, SDValue Op3) { |
| 7681 | SDValue Ops[] = { Op1, Op2, Op3 }; |
| 7682 | return UpdateNodeOperands(N, Ops); |
| 7683 | } |
| 7684 | |
| 7685 | SDNode *SelectionDAG:: |
| 7686 | UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, |
| 7687 | SDValue Op3, SDValue Op4) { |
| 7688 | SDValue Ops[] = { Op1, Op2, Op3, Op4 }; |
| 7689 | return UpdateNodeOperands(N, Ops); |
| 7690 | } |
| 7691 | |
| 7692 | SDNode *SelectionDAG:: |
| 7693 | UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, |
| 7694 | SDValue Op3, SDValue Op4, SDValue Op5) { |
| 7695 | SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 }; |
| 7696 | return UpdateNodeOperands(N, Ops); |
| 7697 | } |
| 7698 | |
| 7699 | SDNode *SelectionDAG:: |
| 7700 | UpdateNodeOperands(SDNode *N, ArrayRef<SDValue> Ops) { |
| 7701 | unsigned NumOps = Ops.size(); |
| 7702 | assert(N->getNumOperands() == NumOps && |
| 7703 | "Update with wrong number of operands" ); |
| 7704 | |
| 7705 | // If no operands changed just return the input node. |
| 7706 | if (std::equal(Ops.begin(), Ops.end(), N->op_begin())) |
| 7707 | return N; |
| 7708 | |
| 7709 | // See if the modified node already exists. |
| 7710 | void *InsertPos = nullptr; |
| 7711 | if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, InsertPos)) |
| 7712 | return Existing; |
| 7713 | |
| 7714 | // Nope it doesn't. Remove the node from its current place in the maps. |
| 7715 | if (InsertPos) |
| 7716 | if (!RemoveNodeFromCSEMaps(N)) |
| 7717 | InsertPos = nullptr; |
| 7718 | |
| 7719 | // Now we update the operands. |
| 7720 | for (unsigned i = 0; i != NumOps; ++i) |
| 7721 | if (N->OperandList[i] != Ops[i]) |
| 7722 | N->OperandList[i].set(Ops[i]); |
| 7723 | |
| 7724 | updateDivergence(N); |
| 7725 | // If this gets put into a CSE map, add it. |
| 7726 | if (InsertPos) CSEMap.InsertNode(N, InsertPos); |
| 7727 | return N; |
| 7728 | } |
| 7729 | |
| 7730 | /// DropOperands - Release the operands and set this node to have |
| 7731 | /// zero operands. |
| 7732 | void SDNode::DropOperands() { |
| 7733 | // Unlike the code in MorphNodeTo that does this, we don't need to |
| 7734 | // watch for dead nodes here. |
| 7735 | for (op_iterator I = op_begin(), E = op_end(); I != E; ) { |
| 7736 | SDUse &Use = *I++; |
| 7737 | Use.set(SDValue()); |
| 7738 | } |
| 7739 | } |
| 7740 | |
| 7741 | void SelectionDAG::setNodeMemRefs(MachineSDNode *N, |
| 7742 | ArrayRef<MachineMemOperand *> NewMemRefs) { |
| 7743 | if (NewMemRefs.empty()) { |
| 7744 | N->clearMemRefs(); |
| 7745 | return; |
| 7746 | } |
| 7747 | |
| 7748 | // Check if we can avoid allocating by storing a single reference directly. |
| 7749 | if (NewMemRefs.size() == 1) { |
| 7750 | N->MemRefs = NewMemRefs[0]; |
| 7751 | N->NumMemRefs = 1; |
| 7752 | return; |
| 7753 | } |
| 7754 | |
| 7755 | MachineMemOperand **MemRefsBuffer = |
| 7756 | Allocator.template Allocate<MachineMemOperand *>(NewMemRefs.size()); |
| 7757 | llvm::copy(NewMemRefs, MemRefsBuffer); |
| 7758 | N->MemRefs = MemRefsBuffer; |
| 7759 | N->NumMemRefs = static_cast<int>(NewMemRefs.size()); |
| 7760 | } |
| 7761 | |
| 7762 | /// SelectNodeTo - These are wrappers around MorphNodeTo that accept a |
| 7763 | /// machine opcode. |
| 7764 | /// |
| 7765 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7766 | EVT VT) { |
| 7767 | SDVTList VTs = getVTList(VT); |
| 7768 | return SelectNodeTo(N, MachineOpc, VTs, None); |
| 7769 | } |
| 7770 | |
| 7771 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7772 | EVT VT, SDValue Op1) { |
| 7773 | SDVTList VTs = getVTList(VT); |
| 7774 | SDValue Ops[] = { Op1 }; |
| 7775 | return SelectNodeTo(N, MachineOpc, VTs, Ops); |
| 7776 | } |
| 7777 | |
| 7778 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7779 | EVT VT, SDValue Op1, |
| 7780 | SDValue Op2) { |
| 7781 | SDVTList VTs = getVTList(VT); |
| 7782 | SDValue Ops[] = { Op1, Op2 }; |
| 7783 | return SelectNodeTo(N, MachineOpc, VTs, Ops); |
| 7784 | } |
| 7785 | |
| 7786 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7787 | EVT VT, SDValue Op1, |
| 7788 | SDValue Op2, SDValue Op3) { |
| 7789 | SDVTList VTs = getVTList(VT); |
| 7790 | SDValue Ops[] = { Op1, Op2, Op3 }; |
| 7791 | return SelectNodeTo(N, MachineOpc, VTs, Ops); |
| 7792 | } |
| 7793 | |
| 7794 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7795 | EVT VT, ArrayRef<SDValue> Ops) { |
| 7796 | SDVTList VTs = getVTList(VT); |
| 7797 | return SelectNodeTo(N, MachineOpc, VTs, Ops); |
| 7798 | } |
| 7799 | |
| 7800 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7801 | EVT VT1, EVT VT2, ArrayRef<SDValue> Ops) { |
| 7802 | SDVTList VTs = getVTList(VT1, VT2); |
| 7803 | return SelectNodeTo(N, MachineOpc, VTs, Ops); |
| 7804 | } |
| 7805 | |
| 7806 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7807 | EVT VT1, EVT VT2) { |
| 7808 | SDVTList VTs = getVTList(VT1, VT2); |
| 7809 | return SelectNodeTo(N, MachineOpc, VTs, None); |
| 7810 | } |
| 7811 | |
| 7812 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7813 | EVT VT1, EVT VT2, EVT VT3, |
| 7814 | ArrayRef<SDValue> Ops) { |
| 7815 | SDVTList VTs = getVTList(VT1, VT2, VT3); |
| 7816 | return SelectNodeTo(N, MachineOpc, VTs, Ops); |
| 7817 | } |
| 7818 | |
| 7819 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7820 | EVT VT1, EVT VT2, |
| 7821 | SDValue Op1, SDValue Op2) { |
| 7822 | SDVTList VTs = getVTList(VT1, VT2); |
| 7823 | SDValue Ops[] = { Op1, Op2 }; |
| 7824 | return SelectNodeTo(N, MachineOpc, VTs, Ops); |
| 7825 | } |
| 7826 | |
| 7827 | SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, |
| 7828 | SDVTList VTs,ArrayRef<SDValue> Ops) { |
| 7829 | SDNode *New = MorphNodeTo(N, ~MachineOpc, VTs, Ops); |
| 7830 | // Reset the NodeID to -1. |
| 7831 | New->setNodeId(-1); |
| 7832 | if (New != N) { |
| 7833 | ReplaceAllUsesWith(N, New); |
| 7834 | RemoveDeadNode(N); |
| 7835 | } |
| 7836 | return New; |
| 7837 | } |
| 7838 | |
| 7839 | /// UpdateSDLocOnMergeSDNode - If the opt level is -O0 then it throws away |
| 7840 | /// the line number information on the merged node since it is not possible to |
| 7841 | /// preserve the information that operation is associated with multiple lines. |
| 7842 | /// This will make the debugger working better at -O0, were there is a higher |
| 7843 | /// probability having other instructions associated with that line. |
| 7844 | /// |
| 7845 | /// For IROrder, we keep the smaller of the two |
| 7846 | SDNode *SelectionDAG::UpdateSDLocOnMergeSDNode(SDNode *N, const SDLoc &OLoc) { |
| 7847 | DebugLoc NLoc = N->getDebugLoc(); |
| 7848 | if (NLoc && OptLevel == CodeGenOpt::None && OLoc.getDebugLoc() != NLoc) { |
| 7849 | N->setDebugLoc(DebugLoc()); |
| 7850 | } |
| 7851 | unsigned Order = std::min(N->getIROrder(), OLoc.getIROrder()); |
| 7852 | N->setIROrder(Order); |
| 7853 | return N; |
| 7854 | } |
| 7855 | |
| 7856 | /// MorphNodeTo - This *mutates* the specified node to have the specified |
| 7857 | /// return type, opcode, and operands. |
| 7858 | /// |
| 7859 | /// Note that MorphNodeTo returns the resultant node. If there is already a |
| 7860 | /// node of the specified opcode and operands, it returns that node instead of |
| 7861 | /// the current one. Note that the SDLoc need not be the same. |
| 7862 | /// |
| 7863 | /// Using MorphNodeTo is faster than creating a new node and swapping it in |
| 7864 | /// with ReplaceAllUsesWith both because it often avoids allocating a new |
| 7865 | /// node, and because it doesn't require CSE recalculation for any of |
| 7866 | /// the node's users. |
| 7867 | /// |
| 7868 | /// However, note that MorphNodeTo recursively deletes dead nodes from the DAG. |
| 7869 | /// As a consequence it isn't appropriate to use from within the DAG combiner or |
| 7870 | /// the legalizer which maintain worklists that would need to be updated when |
| 7871 | /// deleting things. |
| 7872 | SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, |
| 7873 | SDVTList VTs, ArrayRef<SDValue> Ops) { |
| 7874 | // If an identical node already exists, use it. |
| 7875 | void *IP = nullptr; |
| 7876 | if (VTs.VTs[VTs.NumVTs-1] != MVT::Glue) { |
| 7877 | FoldingSetNodeID ID; |
| 7878 | AddNodeIDNode(ID, Opc, VTs, Ops); |
| 7879 | if (SDNode *ON = FindNodeOrInsertPos(ID, SDLoc(N), IP)) |
| 7880 | return UpdateSDLocOnMergeSDNode(ON, SDLoc(N)); |
| 7881 | } |
| 7882 | |
| 7883 | if (!RemoveNodeFromCSEMaps(N)) |
| 7884 | IP = nullptr; |
| 7885 | |
| 7886 | // Start the morphing. |
| 7887 | N->NodeType = Opc; |
| 7888 | N->ValueList = VTs.VTs; |
| 7889 | N->NumValues = VTs.NumVTs; |
| 7890 | |
| 7891 | // Clear the operands list, updating used nodes to remove this from their |
| 7892 | // use list. Keep track of any operands that become dead as a result. |
| 7893 | SmallPtrSet<SDNode*, 16> DeadNodeSet; |
| 7894 | for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) { |
| 7895 | SDUse &Use = *I++; |
| 7896 | SDNode *Used = Use.getNode(); |
| 7897 | Use.set(SDValue()); |
| 7898 | if (Used->use_empty()) |
| 7899 | DeadNodeSet.insert(Used); |
| 7900 | } |
| 7901 | |
| 7902 | // For MachineNode, initialize the memory references information. |
| 7903 | if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) |
| 7904 | MN->clearMemRefs(); |
| 7905 | |
| 7906 | // Swap for an appropriately sized array from the recycler. |
| 7907 | removeOperands(N); |
| 7908 | createOperands(N, Ops); |
| 7909 | |
| 7910 | // Delete any nodes that are still dead after adding the uses for the |
| 7911 | // new operands. |
| 7912 | if (!DeadNodeSet.empty()) { |
| 7913 | SmallVector<SDNode *, 16> DeadNodes; |
| 7914 | for (SDNode *N : DeadNodeSet) |
| 7915 | if (N->use_empty()) |
| 7916 | DeadNodes.push_back(N); |
| 7917 | RemoveDeadNodes(DeadNodes); |
| 7918 | } |
| 7919 | |
| 7920 | if (IP) |
| 7921 | CSEMap.InsertNode(N, IP); // Memoize the new node. |
| 7922 | return N; |
| 7923 | } |
| 7924 | |
| 7925 | SDNode* SelectionDAG::mutateStrictFPToFP(SDNode *Node) { |
| 7926 | unsigned OrigOpc = Node->getOpcode(); |
| 7927 | unsigned NewOpc; |
| 7928 | switch (OrigOpc) { |
| 7929 | default: |
| 7930 | llvm_unreachable("mutateStrictFPToFP called with unexpected opcode!" ); |
| 7931 | case ISD::STRICT_FADD: NewOpc = ISD::FADD; break; |
| 7932 | case ISD::STRICT_FSUB: NewOpc = ISD::FSUB; break; |
| 7933 | case ISD::STRICT_FMUL: NewOpc = ISD::FMUL; break; |
| 7934 | case ISD::STRICT_FDIV: NewOpc = ISD::FDIV; break; |
| 7935 | case ISD::STRICT_FREM: NewOpc = ISD::FREM; break; |
| 7936 | case ISD::STRICT_FMA: NewOpc = ISD::FMA; break; |
| 7937 | case ISD::STRICT_FSQRT: NewOpc = ISD::FSQRT; break; |
| 7938 | case ISD::STRICT_FPOW: NewOpc = ISD::FPOW; break; |
| 7939 | case ISD::STRICT_FPOWI: NewOpc = ISD::FPOWI; break; |
| 7940 | case ISD::STRICT_FSIN: NewOpc = ISD::FSIN; break; |
| 7941 | case ISD::STRICT_FCOS: NewOpc = ISD::FCOS; break; |
| 7942 | case ISD::STRICT_FEXP: NewOpc = ISD::FEXP; break; |
| 7943 | case ISD::STRICT_FEXP2: NewOpc = ISD::FEXP2; break; |
| 7944 | case ISD::STRICT_FLOG: NewOpc = ISD::FLOG; break; |
| 7945 | case ISD::STRICT_FLOG10: NewOpc = ISD::FLOG10; break; |
| 7946 | case ISD::STRICT_FLOG2: NewOpc = ISD::FLOG2; break; |
| 7947 | case ISD::STRICT_FRINT: NewOpc = ISD::FRINT; break; |
| 7948 | case ISD::STRICT_FNEARBYINT: NewOpc = ISD::FNEARBYINT; break; |
| 7949 | case ISD::STRICT_FMAXNUM: NewOpc = ISD::FMAXNUM; break; |
| 7950 | case ISD::STRICT_FMINNUM: NewOpc = ISD::FMINNUM; break; |
| 7951 | case ISD::STRICT_FCEIL: NewOpc = ISD::FCEIL; break; |
| 7952 | case ISD::STRICT_FFLOOR: NewOpc = ISD::FFLOOR; break; |
| 7953 | case ISD::STRICT_FROUND: NewOpc = ISD::FROUND; break; |
| 7954 | case ISD::STRICT_FTRUNC: NewOpc = ISD::FTRUNC; break; |
| 7955 | case ISD::STRICT_FP_ROUND: NewOpc = ISD::FP_ROUND; break; |
| 7956 | case ISD::STRICT_FP_EXTEND: NewOpc = ISD::FP_EXTEND; break; |
| 7957 | } |
| 7958 | |
| 7959 | assert(Node->getNumValues() == 2 && "Unexpected number of results!" ); |
| 7960 | |
| 7961 | // We're taking this node out of the chain, so we need to re-link things. |
| 7962 | SDValue InputChain = Node->getOperand(0); |
| 7963 | SDValue OutputChain = SDValue(Node, 1); |
| 7964 | ReplaceAllUsesOfValueWith(OutputChain, InputChain); |
| 7965 | |
| 7966 | SmallVector<SDValue, 3> Ops; |
| 7967 | for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) |
| 7968 | Ops.push_back(Node->getOperand(i)); |
| 7969 | |
| 7970 | SDVTList VTs = getVTList(Node->getValueType(0)); |
| 7971 | SDNode *Res = MorphNodeTo(Node, NewOpc, VTs, Ops); |
| 7972 | |
| 7973 | // MorphNodeTo can operate in two ways: if an existing node with the |
| 7974 | // specified operands exists, it can just return it. Otherwise, it |
| 7975 | // updates the node in place to have the requested operands. |
| 7976 | if (Res == Node) { |
| 7977 | // If we updated the node in place, reset the node ID. To the isel, |
| 7978 | // this should be just like a newly allocated machine node. |
| 7979 | Res->setNodeId(-1); |
| 7980 | } else { |
| 7981 | ReplaceAllUsesWith(Node, Res); |
| 7982 | RemoveDeadNode(Node); |
| 7983 | } |
| 7984 | |
| 7985 | return Res; |
| 7986 | } |
| 7987 | |
| 7988 | /// getMachineNode - These are used for target selectors to create a new node |
| 7989 | /// with specified return type(s), MachineInstr opcode, and operands. |
| 7990 | /// |
| 7991 | /// Note that getMachineNode returns the resultant node. If there is already a |
| 7992 | /// node of the specified opcode and operands, it returns that node instead of |
| 7993 | /// the current one. |
| 7994 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 7995 | EVT VT) { |
| 7996 | SDVTList VTs = getVTList(VT); |
| 7997 | return getMachineNode(Opcode, dl, VTs, None); |
| 7998 | } |
| 7999 | |
| 8000 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8001 | EVT VT, SDValue Op1) { |
| 8002 | SDVTList VTs = getVTList(VT); |
| 8003 | SDValue Ops[] = { Op1 }; |
| 8004 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8005 | } |
| 8006 | |
| 8007 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8008 | EVT VT, SDValue Op1, SDValue Op2) { |
| 8009 | SDVTList VTs = getVTList(VT); |
| 8010 | SDValue Ops[] = { Op1, Op2 }; |
| 8011 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8012 | } |
| 8013 | |
| 8014 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8015 | EVT VT, SDValue Op1, SDValue Op2, |
| 8016 | SDValue Op3) { |
| 8017 | SDVTList VTs = getVTList(VT); |
| 8018 | SDValue Ops[] = { Op1, Op2, Op3 }; |
| 8019 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8020 | } |
| 8021 | |
| 8022 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8023 | EVT VT, ArrayRef<SDValue> Ops) { |
| 8024 | SDVTList VTs = getVTList(VT); |
| 8025 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8026 | } |
| 8027 | |
| 8028 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8029 | EVT VT1, EVT VT2, SDValue Op1, |
| 8030 | SDValue Op2) { |
| 8031 | SDVTList VTs = getVTList(VT1, VT2); |
| 8032 | SDValue Ops[] = { Op1, Op2 }; |
| 8033 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8034 | } |
| 8035 | |
| 8036 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8037 | EVT VT1, EVT VT2, SDValue Op1, |
| 8038 | SDValue Op2, SDValue Op3) { |
| 8039 | SDVTList VTs = getVTList(VT1, VT2); |
| 8040 | SDValue Ops[] = { Op1, Op2, Op3 }; |
| 8041 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8042 | } |
| 8043 | |
| 8044 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8045 | EVT VT1, EVT VT2, |
| 8046 | ArrayRef<SDValue> Ops) { |
| 8047 | SDVTList VTs = getVTList(VT1, VT2); |
| 8048 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8049 | } |
| 8050 | |
| 8051 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8052 | EVT VT1, EVT VT2, EVT VT3, |
| 8053 | SDValue Op1, SDValue Op2) { |
| 8054 | SDVTList VTs = getVTList(VT1, VT2, VT3); |
| 8055 | SDValue Ops[] = { Op1, Op2 }; |
| 8056 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8057 | } |
| 8058 | |
| 8059 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8060 | EVT VT1, EVT VT2, EVT VT3, |
| 8061 | SDValue Op1, SDValue Op2, |
| 8062 | SDValue Op3) { |
| 8063 | SDVTList VTs = getVTList(VT1, VT2, VT3); |
| 8064 | SDValue Ops[] = { Op1, Op2, Op3 }; |
| 8065 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8066 | } |
| 8067 | |
| 8068 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8069 | EVT VT1, EVT VT2, EVT VT3, |
| 8070 | ArrayRef<SDValue> Ops) { |
| 8071 | SDVTList VTs = getVTList(VT1, VT2, VT3); |
| 8072 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8073 | } |
| 8074 | |
| 8075 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &dl, |
| 8076 | ArrayRef<EVT> ResultTys, |
| 8077 | ArrayRef<SDValue> Ops) { |
| 8078 | SDVTList VTs = getVTList(ResultTys); |
| 8079 | return getMachineNode(Opcode, dl, VTs, Ops); |
| 8080 | } |
| 8081 | |
| 8082 | MachineSDNode *SelectionDAG::getMachineNode(unsigned Opcode, const SDLoc &DL, |
| 8083 | SDVTList VTs, |
| 8084 | ArrayRef<SDValue> Ops) { |
| 8085 | bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Glue; |
| 8086 | MachineSDNode *N; |
| 8087 | void *IP = nullptr; |
| 8088 | |
| 8089 | if (DoCSE) { |
| 8090 | FoldingSetNodeID ID; |
| 8091 | AddNodeIDNode(ID, ~Opcode, VTs, Ops); |
| 8092 | IP = nullptr; |
| 8093 | if (SDNode *E = FindNodeOrInsertPos(ID, DL, IP)) { |
| 8094 | return cast<MachineSDNode>(UpdateSDLocOnMergeSDNode(E, DL)); |
| 8095 | } |
| 8096 | } |
| 8097 | |
| 8098 | // Allocate a new MachineSDNode. |
| 8099 | N = newSDNode<MachineSDNode>(~Opcode, DL.getIROrder(), DL.getDebugLoc(), VTs); |
| 8100 | createOperands(N, Ops); |
| 8101 | |
| 8102 | if (DoCSE) |
| 8103 | CSEMap.InsertNode(N, IP); |
| 8104 | |
| 8105 | InsertNode(N); |
| 8106 | return N; |
| 8107 | } |
| 8108 | |
| 8109 | /// getTargetExtractSubreg - A convenience function for creating |
| 8110 | /// TargetOpcode::EXTRACT_SUBREG nodes. |
| 8111 | SDValue SelectionDAG::(int SRIdx, const SDLoc &DL, EVT VT, |
| 8112 | SDValue Operand) { |
| 8113 | SDValue SRIdxVal = getTargetConstant(SRIdx, DL, MVT::i32); |
| 8114 | SDNode *Subreg = getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, |
| 8115 | VT, Operand, SRIdxVal); |
| 8116 | return SDValue(Subreg, 0); |
| 8117 | } |
| 8118 | |
| 8119 | /// getTargetInsertSubreg - A convenience function for creating |
| 8120 | /// TargetOpcode::INSERT_SUBREG nodes. |
| 8121 | SDValue SelectionDAG::getTargetInsertSubreg(int SRIdx, const SDLoc &DL, EVT VT, |
| 8122 | SDValue Operand, SDValue Subreg) { |
| 8123 | SDValue SRIdxVal = getTargetConstant(SRIdx, DL, MVT::i32); |
| 8124 | SDNode *Result = getMachineNode(TargetOpcode::INSERT_SUBREG, DL, |
| 8125 | VT, Operand, Subreg, SRIdxVal); |
| 8126 | return SDValue(Result, 0); |
| 8127 | } |
| 8128 | |
| 8129 | /// getNodeIfExists - Get the specified node if it's already available, or |
| 8130 | /// else return NULL. |
| 8131 | SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList, |
| 8132 | ArrayRef<SDValue> Ops, |
| 8133 | const SDNodeFlags Flags) { |
| 8134 | if (VTList.VTs[VTList.NumVTs - 1] != MVT::Glue) { |
| 8135 | FoldingSetNodeID ID; |
| 8136 | AddNodeIDNode(ID, Opcode, VTList, Ops); |
| 8137 | void *IP = nullptr; |
| 8138 | if (SDNode *E = FindNodeOrInsertPos(ID, SDLoc(), IP)) { |
| 8139 | E->intersectFlagsWith(Flags); |
| 8140 | return E; |
| 8141 | } |
| 8142 | } |
| 8143 | return nullptr; |
| 8144 | } |
| 8145 | |
| 8146 | /// getDbgValue - Creates a SDDbgValue node. |
| 8147 | /// |
| 8148 | /// SDNode |
| 8149 | SDDbgValue *SelectionDAG::getDbgValue(DIVariable *Var, DIExpression *Expr, |
| 8150 | SDNode *N, unsigned R, bool IsIndirect, |
| 8151 | const DebugLoc &DL, unsigned O) { |
| 8152 | assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) && |
| 8153 | "Expected inlined-at fields to agree" ); |
| 8154 | return new (DbgInfo->getAlloc()) |
| 8155 | SDDbgValue(Var, Expr, N, R, IsIndirect, DL, O); |
| 8156 | } |
| 8157 | |
| 8158 | /// Constant |
| 8159 | SDDbgValue *SelectionDAG::getConstantDbgValue(DIVariable *Var, |
| 8160 | DIExpression *Expr, |
| 8161 | const Value *C, |
| 8162 | const DebugLoc &DL, unsigned O) { |
| 8163 | assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) && |
| 8164 | "Expected inlined-at fields to agree" ); |
| 8165 | return new (DbgInfo->getAlloc()) SDDbgValue(Var, Expr, C, DL, O); |
| 8166 | } |
| 8167 | |
| 8168 | /// FrameIndex |
| 8169 | SDDbgValue *SelectionDAG::getFrameIndexDbgValue(DIVariable *Var, |
| 8170 | DIExpression *Expr, unsigned FI, |
| 8171 | bool IsIndirect, |
| 8172 | const DebugLoc &DL, |
| 8173 | unsigned O) { |
| 8174 | assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) && |
| 8175 | "Expected inlined-at fields to agree" ); |
| 8176 | return new (DbgInfo->getAlloc()) |
| 8177 | SDDbgValue(Var, Expr, FI, IsIndirect, DL, O, SDDbgValue::FRAMEIX); |
| 8178 | } |
| 8179 | |
| 8180 | /// VReg |
| 8181 | SDDbgValue *SelectionDAG::getVRegDbgValue(DIVariable *Var, |
| 8182 | DIExpression *Expr, |
| 8183 | unsigned VReg, bool IsIndirect, |
| 8184 | const DebugLoc &DL, unsigned O) { |
| 8185 | assert(cast<DILocalVariable>(Var)->isValidLocationForIntrinsic(DL) && |
| 8186 | "Expected inlined-at fields to agree" ); |
| 8187 | return new (DbgInfo->getAlloc()) |
| 8188 | SDDbgValue(Var, Expr, VReg, IsIndirect, DL, O, SDDbgValue::VREG); |
| 8189 | } |
| 8190 | |
| 8191 | void SelectionDAG::transferDbgValues(SDValue From, SDValue To, |
| 8192 | unsigned OffsetInBits, unsigned SizeInBits, |
| 8193 | bool InvalidateDbg) { |
| 8194 | SDNode *FromNode = From.getNode(); |
| 8195 | SDNode *ToNode = To.getNode(); |
| 8196 | assert(FromNode && ToNode && "Can't modify dbg values" ); |
| 8197 | |
| 8198 | // PR35338 |
| 8199 | // TODO: assert(From != To && "Redundant dbg value transfer"); |
| 8200 | // TODO: assert(FromNode != ToNode && "Intranode dbg value transfer"); |
| 8201 | if (From == To || FromNode == ToNode) |
| 8202 | return; |
| 8203 | |
| 8204 | if (!FromNode->getHasDebugValue()) |
| 8205 | return; |
| 8206 | |
| 8207 | SmallVector<SDDbgValue *, 2> ClonedDVs; |
| 8208 | for (SDDbgValue *Dbg : GetDbgValues(FromNode)) { |
| 8209 | if (Dbg->getKind() != SDDbgValue::SDNODE || Dbg->isInvalidated()) |
| 8210 | continue; |
| 8211 | |
| 8212 | // TODO: assert(!Dbg->isInvalidated() && "Transfer of invalid dbg value"); |
| 8213 | |
| 8214 | // Just transfer the dbg value attached to From. |
| 8215 | if (Dbg->getResNo() != From.getResNo()) |
| 8216 | continue; |
| 8217 | |
| 8218 | DIVariable *Var = Dbg->getVariable(); |
| 8219 | auto *Expr = Dbg->getExpression(); |
| 8220 | // If a fragment is requested, update the expression. |
| 8221 | if (SizeInBits) { |
| 8222 | // When splitting a larger (e.g., sign-extended) value whose |
| 8223 | // lower bits are described with an SDDbgValue, do not attempt |
| 8224 | // to transfer the SDDbgValue to the upper bits. |
| 8225 | if (auto FI = Expr->getFragmentInfo()) |
| 8226 | if (OffsetInBits + SizeInBits > FI->SizeInBits) |
| 8227 | continue; |
| 8228 | auto Fragment = DIExpression::createFragmentExpression(Expr, OffsetInBits, |
| 8229 | SizeInBits); |
| 8230 | if (!Fragment) |
| 8231 | continue; |
| 8232 | Expr = *Fragment; |
| 8233 | } |
| 8234 | // Clone the SDDbgValue and move it to To. |
| 8235 | SDDbgValue *Clone = |
| 8236 | getDbgValue(Var, Expr, ToNode, To.getResNo(), Dbg->isIndirect(), |
| 8237 | Dbg->getDebugLoc(), Dbg->getOrder()); |
| 8238 | ClonedDVs.push_back(Clone); |
| 8239 | |
| 8240 | if (InvalidateDbg) { |
| 8241 | // Invalidate value and indicate the SDDbgValue should not be emitted. |
| 8242 | Dbg->setIsInvalidated(); |
| 8243 | Dbg->setIsEmitted(); |
| 8244 | } |
| 8245 | } |
| 8246 | |
| 8247 | for (SDDbgValue *Dbg : ClonedDVs) |
| 8248 | AddDbgValue(Dbg, ToNode, false); |
| 8249 | } |
| 8250 | |
| 8251 | void SelectionDAG::salvageDebugInfo(SDNode &N) { |
| 8252 | if (!N.getHasDebugValue()) |
| 8253 | return; |
| 8254 | |
| 8255 | SmallVector<SDDbgValue *, 2> ClonedDVs; |
| 8256 | for (auto DV : GetDbgValues(&N)) { |
| 8257 | if (DV->isInvalidated()) |
| 8258 | continue; |
| 8259 | switch (N.getOpcode()) { |
| 8260 | default: |
| 8261 | break; |
| 8262 | case ISD::ADD: |
| 8263 | SDValue N0 = N.getOperand(0); |
| 8264 | SDValue N1 = N.getOperand(1); |
| 8265 | if (!isConstantIntBuildVectorOrConstantInt(N0) && |
| 8266 | isConstantIntBuildVectorOrConstantInt(N1)) { |
| 8267 | uint64_t Offset = N.getConstantOperandVal(1); |
| 8268 | // Rewrite an ADD constant node into a DIExpression. Since we are |
| 8269 | // performing arithmetic to compute the variable's *value* in the |
| 8270 | // DIExpression, we need to mark the expression with a |
| 8271 | // DW_OP_stack_value. |
| 8272 | auto *DIExpr = DV->getExpression(); |
| 8273 | DIExpr = |
| 8274 | DIExpression::prepend(DIExpr, DIExpression::StackValue, Offset); |
| 8275 | SDDbgValue *Clone = |
| 8276 | getDbgValue(DV->getVariable(), DIExpr, N0.getNode(), N0.getResNo(), |
| 8277 | DV->isIndirect(), DV->getDebugLoc(), DV->getOrder()); |
| 8278 | ClonedDVs.push_back(Clone); |
| 8279 | DV->setIsInvalidated(); |
| 8280 | DV->setIsEmitted(); |
| 8281 | LLVM_DEBUG(dbgs() << "SALVAGE: Rewriting" ; |
| 8282 | N0.getNode()->dumprFull(this); |
| 8283 | dbgs() << " into " << *DIExpr << '\n'); |
| 8284 | } |
| 8285 | } |
| 8286 | } |
| 8287 | |
| 8288 | for (SDDbgValue *Dbg : ClonedDVs) |
| 8289 | AddDbgValue(Dbg, Dbg->getSDNode(), false); |
| 8290 | } |
| 8291 | |
| 8292 | /// Creates a SDDbgLabel node. |
| 8293 | SDDbgLabel *SelectionDAG::getDbgLabel(DILabel *Label, |
| 8294 | const DebugLoc &DL, unsigned O) { |
| 8295 | assert(cast<DILabel>(Label)->isValidLocationForIntrinsic(DL) && |
| 8296 | "Expected inlined-at fields to agree" ); |
| 8297 | return new (DbgInfo->getAlloc()) SDDbgLabel(Label, DL, O); |
| 8298 | } |
| 8299 | |
| 8300 | namespace { |
| 8301 | |
| 8302 | /// RAUWUpdateListener - Helper for ReplaceAllUsesWith - When the node |
| 8303 | /// pointed to by a use iterator is deleted, increment the use iterator |
| 8304 | /// so that it doesn't dangle. |
| 8305 | /// |
| 8306 | class RAUWUpdateListener : public SelectionDAG::DAGUpdateListener { |
| 8307 | SDNode::use_iterator &UI; |
| 8308 | SDNode::use_iterator &UE; |
| 8309 | |
| 8310 | void NodeDeleted(SDNode *N, SDNode *E) override { |
| 8311 | // Increment the iterator as needed. |
| 8312 | while (UI != UE && N == *UI) |
| 8313 | ++UI; |
| 8314 | } |
| 8315 | |
| 8316 | public: |
| 8317 | RAUWUpdateListener(SelectionDAG &d, |
| 8318 | SDNode::use_iterator &ui, |
| 8319 | SDNode::use_iterator &ue) |
| 8320 | : SelectionDAG::DAGUpdateListener(d), UI(ui), UE(ue) {} |
| 8321 | }; |
| 8322 | |
| 8323 | } // end anonymous namespace |
| 8324 | |
| 8325 | /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. |
| 8326 | /// This can cause recursive merging of nodes in the DAG. |
| 8327 | /// |
| 8328 | /// This version assumes From has a single result value. |
| 8329 | /// |
| 8330 | void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To) { |
| 8331 | SDNode *From = FromN.getNode(); |
| 8332 | assert(From->getNumValues() == 1 && FromN.getResNo() == 0 && |
| 8333 | "Cannot replace with this method!" ); |
| 8334 | assert(From != To.getNode() && "Cannot replace uses of with self" ); |
| 8335 | |
| 8336 | // Preserve Debug Values |
| 8337 | transferDbgValues(FromN, To); |
| 8338 | |
| 8339 | // Iterate over all the existing uses of From. New uses will be added |
| 8340 | // to the beginning of the use list, which we avoid visiting. |
| 8341 | // This specifically avoids visiting uses of From that arise while the |
| 8342 | // replacement is happening, because any such uses would be the result |
| 8343 | // of CSE: If an existing node looks like From after one of its operands |
| 8344 | // is replaced by To, we don't want to replace of all its users with To |
| 8345 | // too. See PR3018 for more info. |
| 8346 | SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); |
| 8347 | RAUWUpdateListener Listener(*this, UI, UE); |
| 8348 | while (UI != UE) { |
| 8349 | SDNode *User = *UI; |
| 8350 | |
| 8351 | // This node is about to morph, remove its old self from the CSE maps. |
| 8352 | RemoveNodeFromCSEMaps(User); |
| 8353 | |
| 8354 | // A user can appear in a use list multiple times, and when this |
| 8355 | // happens the uses are usually next to each other in the list. |
| 8356 | // To help reduce the number of CSE recomputations, process all |
| 8357 | // the uses of this user that we can find this way. |
| 8358 | do { |
| 8359 | SDUse &Use = UI.getUse(); |
| 8360 | ++UI; |
| 8361 | Use.set(To); |
| 8362 | if (To->isDivergent() != From->isDivergent()) |
| 8363 | updateDivergence(User); |
| 8364 | } while (UI != UE && *UI == User); |
| 8365 | // Now that we have modified User, add it back to the CSE maps. If it |
| 8366 | // already exists there, recursively merge the results together. |
| 8367 | AddModifiedNodeToCSEMaps(User); |
| 8368 | } |
| 8369 | |
| 8370 | // If we just RAUW'd the root, take note. |
| 8371 | if (FromN == getRoot()) |
| 8372 | setRoot(To); |
| 8373 | } |
| 8374 | |
| 8375 | /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. |
| 8376 | /// This can cause recursive merging of nodes in the DAG. |
| 8377 | /// |
| 8378 | /// This version assumes that for each value of From, there is a |
| 8379 | /// corresponding value in To in the same position with the same type. |
| 8380 | /// |
| 8381 | void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To) { |
| 8382 | #ifndef NDEBUG |
| 8383 | for (unsigned i = 0, e = From->getNumValues(); i != e; ++i) |
| 8384 | assert((!From->hasAnyUseOfValue(i) || |
| 8385 | From->getValueType(i) == To->getValueType(i)) && |
| 8386 | "Cannot use this version of ReplaceAllUsesWith!" ); |
| 8387 | #endif |
| 8388 | |
| 8389 | // Handle the trivial case. |
| 8390 | if (From == To) |
| 8391 | return; |
| 8392 | |
| 8393 | // Preserve Debug Info. Only do this if there's a use. |
| 8394 | for (unsigned i = 0, e = From->getNumValues(); i != e; ++i) |
| 8395 | if (From->hasAnyUseOfValue(i)) { |
| 8396 | assert((i < To->getNumValues()) && "Invalid To location" ); |
| 8397 | transferDbgValues(SDValue(From, i), SDValue(To, i)); |
| 8398 | } |
| 8399 | |
| 8400 | // Iterate over just the existing users of From. See the comments in |
| 8401 | // the ReplaceAllUsesWith above. |
| 8402 | SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); |
| 8403 | RAUWUpdateListener Listener(*this, UI, UE); |
| 8404 | while (UI != UE) { |
| 8405 | SDNode *User = *UI; |
| 8406 | |
| 8407 | // This node is about to morph, remove its old self from the CSE maps. |
| 8408 | RemoveNodeFromCSEMaps(User); |
| 8409 | |
| 8410 | // A user can appear in a use list multiple times, and when this |
| 8411 | // happens the uses are usually next to each other in the list. |
| 8412 | // To help reduce the number of CSE recomputations, process all |
| 8413 | // the uses of this user that we can find this way. |
| 8414 | do { |
| 8415 | SDUse &Use = UI.getUse(); |
| 8416 | ++UI; |
| 8417 | Use.setNode(To); |
| 8418 | if (To->isDivergent() != From->isDivergent()) |
| 8419 | updateDivergence(User); |
| 8420 | } while (UI != UE && *UI == User); |
| 8421 | |
| 8422 | // Now that we have modified User, add it back to the CSE maps. If it |
| 8423 | // already exists there, recursively merge the results together. |
| 8424 | AddModifiedNodeToCSEMaps(User); |
| 8425 | } |
| 8426 | |
| 8427 | // If we just RAUW'd the root, take note. |
| 8428 | if (From == getRoot().getNode()) |
| 8429 | setRoot(SDValue(To, getRoot().getResNo())); |
| 8430 | } |
| 8431 | |
| 8432 | /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. |
| 8433 | /// This can cause recursive merging of nodes in the DAG. |
| 8434 | /// |
| 8435 | /// This version can replace From with any result values. To must match the |
| 8436 | /// number and types of values returned by From. |
| 8437 | void SelectionDAG::ReplaceAllUsesWith(SDNode *From, const SDValue *To) { |
| 8438 | if (From->getNumValues() == 1) // Handle the simple case efficiently. |
| 8439 | return ReplaceAllUsesWith(SDValue(From, 0), To[0]); |
| 8440 | |
| 8441 | // Preserve Debug Info. |
| 8442 | for (unsigned i = 0, e = From->getNumValues(); i != e; ++i) |
| 8443 | transferDbgValues(SDValue(From, i), To[i]); |
| 8444 | |
| 8445 | // Iterate over just the existing users of From. See the comments in |
| 8446 | // the ReplaceAllUsesWith above. |
| 8447 | SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); |
| 8448 | RAUWUpdateListener Listener(*this, UI, UE); |
| 8449 | while (UI != UE) { |
| 8450 | SDNode *User = *UI; |
| 8451 | |
| 8452 | // This node is about to morph, remove its old self from the CSE maps. |
| 8453 | RemoveNodeFromCSEMaps(User); |
| 8454 | |
| 8455 | // A user can appear in a use list multiple times, and when this happens the |
| 8456 | // uses are usually next to each other in the list. To help reduce the |
| 8457 | // number of CSE and divergence recomputations, process all the uses of this |
| 8458 | // user that we can find this way. |
| 8459 | bool To_IsDivergent = false; |
| 8460 | do { |
| 8461 | SDUse &Use = UI.getUse(); |
| 8462 | const SDValue &ToOp = To[Use.getResNo()]; |
| 8463 | ++UI; |
| 8464 | Use.set(ToOp); |
| 8465 | To_IsDivergent |= ToOp->isDivergent(); |
| 8466 | } while (UI != UE && *UI == User); |
| 8467 | |
| 8468 | if (To_IsDivergent != From->isDivergent()) |
| 8469 | updateDivergence(User); |
| 8470 | |
| 8471 | // Now that we have modified User, add it back to the CSE maps. If it |
| 8472 | // already exists there, recursively merge the results together. |
| 8473 | AddModifiedNodeToCSEMaps(User); |
| 8474 | } |
| 8475 | |
| 8476 | // If we just RAUW'd the root, take note. |
| 8477 | if (From == getRoot().getNode()) |
| 8478 | setRoot(SDValue(To[getRoot().getResNo()])); |
| 8479 | } |
| 8480 | |
| 8481 | /// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving |
| 8482 | /// uses of other values produced by From.getNode() alone. The Deleted |
| 8483 | /// vector is handled the same way as for ReplaceAllUsesWith. |
| 8484 | void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To){ |
| 8485 | // Handle the really simple, really trivial case efficiently. |
| 8486 | if (From == To) return; |
| 8487 | |
| 8488 | // Handle the simple, trivial, case efficiently. |
| 8489 | if (From.getNode()->getNumValues() == 1) { |
| 8490 | ReplaceAllUsesWith(From, To); |
| 8491 | return; |
| 8492 | } |
| 8493 | |
| 8494 | // Preserve Debug Info. |
| 8495 | transferDbgValues(From, To); |
| 8496 | |
| 8497 | // Iterate over just the existing users of From. See the comments in |
| 8498 | // the ReplaceAllUsesWith above. |
| 8499 | SDNode::use_iterator UI = From.getNode()->use_begin(), |
| 8500 | UE = From.getNode()->use_end(); |
| 8501 | RAUWUpdateListener Listener(*this, UI, UE); |
| 8502 | while (UI != UE) { |
| 8503 | SDNode *User = *UI; |
| 8504 | bool UserRemovedFromCSEMaps = false; |
| 8505 | |
| 8506 | // A user can appear in a use list multiple times, and when this |
| 8507 | // happens the uses are usually next to each other in the list. |
| 8508 | // To help reduce the number of CSE recomputations, process all |
| 8509 | // the uses of this user that we can find this way. |
| 8510 | do { |
| 8511 | SDUse &Use = UI.getUse(); |
| 8512 | |
| 8513 | // Skip uses of different values from the same node. |
| 8514 | if (Use.getResNo() != From.getResNo()) { |
| 8515 | ++UI; |
| 8516 | continue; |
| 8517 | } |
| 8518 | |
| 8519 | // If this node hasn't been modified yet, it's still in the CSE maps, |
| 8520 | // so remove its old self from the CSE maps. |
| 8521 | if (!UserRemovedFromCSEMaps) { |
| 8522 | RemoveNodeFromCSEMaps(User); |
| 8523 | UserRemovedFromCSEMaps = true; |
| 8524 | } |
| 8525 | |
| 8526 | ++UI; |
| 8527 | Use.set(To); |
| 8528 | if (To->isDivergent() != From->isDivergent()) |
| 8529 | updateDivergence(User); |
| 8530 | } while (UI != UE && *UI == User); |
| 8531 | // We are iterating over all uses of the From node, so if a use |
| 8532 | // doesn't use the specific value, no changes are made. |
| 8533 | if (!UserRemovedFromCSEMaps) |
| 8534 | continue; |
| 8535 | |
| 8536 | // Now that we have modified User, add it back to the CSE maps. If it |
| 8537 | // already exists there, recursively merge the results together. |
| 8538 | AddModifiedNodeToCSEMaps(User); |
| 8539 | } |
| 8540 | |
| 8541 | // If we just RAUW'd the root, take note. |
| 8542 | if (From == getRoot()) |
| 8543 | setRoot(To); |
| 8544 | } |
| 8545 | |
| 8546 | namespace { |
| 8547 | |
| 8548 | /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith |
| 8549 | /// to record information about a use. |
| 8550 | struct UseMemo { |
| 8551 | SDNode *User; |
| 8552 | unsigned Index; |
| 8553 | SDUse *Use; |
| 8554 | }; |
| 8555 | |
| 8556 | /// operator< - Sort Memos by User. |
| 8557 | bool operator<(const UseMemo &L, const UseMemo &R) { |
| 8558 | return (intptr_t)L.User < (intptr_t)R.User; |
| 8559 | } |
| 8560 | |
| 8561 | } // end anonymous namespace |
| 8562 | |
| 8563 | void SelectionDAG::updateDivergence(SDNode * N) |
| 8564 | { |
| 8565 | if (TLI->isSDNodeAlwaysUniform(N)) |
| 8566 | return; |
| 8567 | bool IsDivergent = TLI->isSDNodeSourceOfDivergence(N, FLI, DA); |
| 8568 | for (auto &Op : N->ops()) { |
| 8569 | if (Op.Val.getValueType() != MVT::Other) |
| 8570 | IsDivergent |= Op.getNode()->isDivergent(); |
| 8571 | } |
| 8572 | if (N->SDNodeBits.IsDivergent != IsDivergent) { |
| 8573 | N->SDNodeBits.IsDivergent = IsDivergent; |
| 8574 | for (auto U : N->uses()) { |
| 8575 | updateDivergence(U); |
| 8576 | } |
| 8577 | } |
| 8578 | } |
| 8579 | |
| 8580 | void SelectionDAG::CreateTopologicalOrder(std::vector<SDNode *> &Order) { |
| 8581 | DenseMap<SDNode *, unsigned> Degree; |
| 8582 | Order.reserve(AllNodes.size()); |
| 8583 | for (auto &N : allnodes()) { |
| 8584 | unsigned NOps = N.getNumOperands(); |
| 8585 | Degree[&N] = NOps; |
| 8586 | if (0 == NOps) |
| 8587 | Order.push_back(&N); |
| 8588 | } |
| 8589 | for (size_t I = 0; I != Order.size(); ++I) { |
| 8590 | SDNode *N = Order[I]; |
| 8591 | for (auto U : N->uses()) { |
| 8592 | unsigned &UnsortedOps = Degree[U]; |
| 8593 | if (0 == --UnsortedOps) |
| 8594 | Order.push_back(U); |
| 8595 | } |
| 8596 | } |
| 8597 | } |
| 8598 | |
| 8599 | #ifndef NDEBUG |
| 8600 | void SelectionDAG::VerifyDAGDiverence() { |
| 8601 | std::vector<SDNode *> TopoOrder; |
| 8602 | CreateTopologicalOrder(TopoOrder); |
| 8603 | const TargetLowering &TLI = getTargetLoweringInfo(); |
| 8604 | DenseMap<const SDNode *, bool> DivergenceMap; |
| 8605 | for (auto &N : allnodes()) { |
| 8606 | DivergenceMap[&N] = false; |
| 8607 | } |
| 8608 | for (auto N : TopoOrder) { |
| 8609 | bool IsDivergent = DivergenceMap[N]; |
| 8610 | bool IsSDNodeDivergent = TLI.isSDNodeSourceOfDivergence(N, FLI, DA); |
| 8611 | for (auto &Op : N->ops()) { |
| 8612 | if (Op.Val.getValueType() != MVT::Other) |
| 8613 | IsSDNodeDivergent |= DivergenceMap[Op.getNode()]; |
| 8614 | } |
| 8615 | if (!IsDivergent && IsSDNodeDivergent && !TLI.isSDNodeAlwaysUniform(N)) { |
| 8616 | DivergenceMap[N] = true; |
| 8617 | } |
| 8618 | } |
| 8619 | for (auto &N : allnodes()) { |
| 8620 | (void)N; |
| 8621 | assert(DivergenceMap[&N] == N.isDivergent() && |
| 8622 | "Divergence bit inconsistency detected\n" ); |
| 8623 | } |
| 8624 | } |
| 8625 | #endif |
| 8626 | |
| 8627 | /// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving |
| 8628 | /// uses of other values produced by From.getNode() alone. The same value |
| 8629 | /// may appear in both the From and To list. The Deleted vector is |
| 8630 | /// handled the same way as for ReplaceAllUsesWith. |
| 8631 | void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From, |
| 8632 | const SDValue *To, |
| 8633 | unsigned Num){ |
| 8634 | // Handle the simple, trivial case efficiently. |
| 8635 | if (Num == 1) |
| 8636 | return ReplaceAllUsesOfValueWith(*From, *To); |
| 8637 | |
| 8638 | transferDbgValues(*From, *To); |
| 8639 | |
| 8640 | // Read up all the uses and make records of them. This helps |
| 8641 | // processing new uses that are introduced during the |
| 8642 | // replacement process. |
| 8643 | SmallVector<UseMemo, 4> Uses; |
| 8644 | for (unsigned i = 0; i != Num; ++i) { |
| 8645 | unsigned FromResNo = From[i].getResNo(); |
| 8646 | SDNode *FromNode = From[i].getNode(); |
| 8647 | for (SDNode::use_iterator UI = FromNode->use_begin(), |
| 8648 | E = FromNode->use_end(); UI != E; ++UI) { |
| 8649 | SDUse &Use = UI.getUse(); |
| 8650 | if (Use.getResNo() == FromResNo) { |
| 8651 | UseMemo Memo = { *UI, i, &Use }; |
| 8652 | Uses.push_back(Memo); |
| 8653 | } |
| 8654 | } |
| 8655 | } |
| 8656 | |
| 8657 | // Sort the uses, so that all the uses from a given User are together. |
| 8658 | llvm::sort(Uses); |
| 8659 | |
| 8660 | for (unsigned UseIndex = 0, UseIndexEnd = Uses.size(); |
| 8661 | UseIndex != UseIndexEnd; ) { |
| 8662 | // We know that this user uses some value of From. If it is the right |
| 8663 | // value, update it. |
| 8664 | SDNode *User = Uses[UseIndex].User; |
| 8665 | |
| 8666 | // This node is about to morph, remove its old self from the CSE maps. |
| 8667 | RemoveNodeFromCSEMaps(User); |
| 8668 | |
| 8669 | // The Uses array is sorted, so all the uses for a given User |
| 8670 | // are next to each other in the list. |
| 8671 | // To help reduce the number of CSE recomputations, process all |
| 8672 | // the uses of this user that we can find this way. |
| 8673 | do { |
| 8674 | unsigned i = Uses[UseIndex].Index; |
| 8675 | SDUse &Use = *Uses[UseIndex].Use; |
| 8676 | ++UseIndex; |
| 8677 | |
| 8678 | Use.set(To[i]); |
| 8679 | } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User); |
| 8680 | |
| 8681 | // Now that we have modified User, add it back to the CSE maps. If it |
| 8682 | // already exists there, recursively merge the results together. |
| 8683 | AddModifiedNodeToCSEMaps(User); |
| 8684 | } |
| 8685 | } |
| 8686 | |
| 8687 | /// AssignTopologicalOrder - Assign a unique node id for each node in the DAG |
| 8688 | /// based on their topological order. It returns the maximum id and a vector |
| 8689 | /// of the SDNodes* in assigned order by reference. |
| 8690 | unsigned SelectionDAG::AssignTopologicalOrder() { |
| 8691 | unsigned DAGSize = 0; |
| 8692 | |
| 8693 | // SortedPos tracks the progress of the algorithm. Nodes before it are |
| 8694 | // sorted, nodes after it are unsorted. When the algorithm completes |
| 8695 | // it is at the end of the list. |
| 8696 | allnodes_iterator SortedPos = allnodes_begin(); |
| 8697 | |
| 8698 | // Visit all the nodes. Move nodes with no operands to the front of |
| 8699 | // the list immediately. Annotate nodes that do have operands with their |
| 8700 | // operand count. Before we do this, the Node Id fields of the nodes |
| 8701 | // may contain arbitrary values. After, the Node Id fields for nodes |
| 8702 | // before SortedPos will contain the topological sort index, and the |
| 8703 | // Node Id fields for nodes At SortedPos and after will contain the |
| 8704 | // count of outstanding operands. |
| 8705 | for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) { |
| 8706 | SDNode *N = &*I++; |
| 8707 | checkForCycles(N, this); |
| 8708 | unsigned Degree = N->getNumOperands(); |
| 8709 | if (Degree == 0) { |
| 8710 | // A node with no uses, add it to the result array immediately. |
| 8711 | N->setNodeId(DAGSize++); |
| 8712 | allnodes_iterator Q(N); |
| 8713 | if (Q != SortedPos) |
| 8714 | SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q)); |
| 8715 | assert(SortedPos != AllNodes.end() && "Overran node list" ); |
| 8716 | ++SortedPos; |
| 8717 | } else { |
| 8718 | // Temporarily use the Node Id as scratch space for the degree count. |
| 8719 | N->setNodeId(Degree); |
| 8720 | } |
| 8721 | } |
| 8722 | |
| 8723 | // Visit all the nodes. As we iterate, move nodes into sorted order, |
| 8724 | // such that by the time the end is reached all nodes will be sorted. |
| 8725 | for (SDNode &Node : allnodes()) { |
| 8726 | SDNode *N = &Node; |
| 8727 | checkForCycles(N, this); |
| 8728 | // N is in sorted position, so all its uses have one less operand |
| 8729 | // that needs to be sorted. |
| 8730 | for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); |
| 8731 | UI != UE; ++UI) { |
| 8732 | SDNode *P = *UI; |
| 8733 | unsigned Degree = P->getNodeId(); |
| 8734 | assert(Degree != 0 && "Invalid node degree" ); |
| 8735 | --Degree; |
| 8736 | if (Degree == 0) { |
| 8737 | // All of P's operands are sorted, so P may sorted now. |
| 8738 | P->setNodeId(DAGSize++); |
| 8739 | if (P->getIterator() != SortedPos) |
| 8740 | SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P)); |
| 8741 | assert(SortedPos != AllNodes.end() && "Overran node list" ); |
| 8742 | ++SortedPos; |
| 8743 | } else { |
| 8744 | // Update P's outstanding operand count. |
| 8745 | P->setNodeId(Degree); |
| 8746 | } |
| 8747 | } |
| 8748 | if (Node.getIterator() == SortedPos) { |
| 8749 | #ifndef NDEBUG |
| 8750 | allnodes_iterator I(N); |
| 8751 | SDNode *S = &*++I; |
| 8752 | dbgs() << "Overran sorted position:\n" ; |
| 8753 | S->dumprFull(this); dbgs() << "\n" ; |
| 8754 | dbgs() << "Checking if this is due to cycles\n" ; |
| 8755 | checkForCycles(this, true); |
| 8756 | #endif |
| 8757 | llvm_unreachable(nullptr); |
| 8758 | } |
| 8759 | } |
| 8760 | |
| 8761 | assert(SortedPos == AllNodes.end() && |
| 8762 | "Topological sort incomplete!" ); |
| 8763 | assert(AllNodes.front().getOpcode() == ISD::EntryToken && |
| 8764 | "First node in topological sort is not the entry token!" ); |
| 8765 | assert(AllNodes.front().getNodeId() == 0 && |
| 8766 | "First node in topological sort has non-zero id!" ); |
| 8767 | assert(AllNodes.front().getNumOperands() == 0 && |
| 8768 | "First node in topological sort has operands!" ); |
| 8769 | assert(AllNodes.back().getNodeId() == (int)DAGSize-1 && |
| 8770 | "Last node in topologic sort has unexpected id!" ); |
| 8771 | assert(AllNodes.back().use_empty() && |
| 8772 | "Last node in topologic sort has users!" ); |
| 8773 | assert(DAGSize == allnodes_size() && "Node count mismatch!" ); |
| 8774 | return DAGSize; |
| 8775 | } |
| 8776 | |
| 8777 | /// AddDbgValue - Add a dbg_value SDNode. If SD is non-null that means the |
| 8778 | /// value is produced by SD. |
| 8779 | void SelectionDAG::AddDbgValue(SDDbgValue *DB, SDNode *SD, bool isParameter) { |
| 8780 | if (SD) { |
| 8781 | assert(DbgInfo->getSDDbgValues(SD).empty() || SD->getHasDebugValue()); |
| 8782 | SD->setHasDebugValue(true); |
| 8783 | } |
| 8784 | DbgInfo->add(DB, SD, isParameter); |
| 8785 | } |
| 8786 | |
| 8787 | void SelectionDAG::AddDbgLabel(SDDbgLabel *DB) { |
| 8788 | DbgInfo->add(DB); |
| 8789 | } |
| 8790 | |
| 8791 | SDValue SelectionDAG::makeEquivalentMemoryOrdering(LoadSDNode *OldLoad, |
| 8792 | SDValue NewMemOp) { |
| 8793 | assert(isa<MemSDNode>(NewMemOp.getNode()) && "Expected a memop node" ); |
| 8794 | // The new memory operation must have the same position as the old load in |
| 8795 | // terms of memory dependency. Create a TokenFactor for the old load and new |
| 8796 | // memory operation and update uses of the old load's output chain to use that |
| 8797 | // TokenFactor. |
| 8798 | SDValue OldChain = SDValue(OldLoad, 1); |
| 8799 | SDValue NewChain = SDValue(NewMemOp.getNode(), 1); |
| 8800 | if (!OldLoad->hasAnyUseOfValue(1)) |
| 8801 | return NewChain; |
| 8802 | |
| 8803 | SDValue TokenFactor = |
| 8804 | getNode(ISD::TokenFactor, SDLoc(OldLoad), MVT::Other, OldChain, NewChain); |
| 8805 | ReplaceAllUsesOfValueWith(OldChain, TokenFactor); |
| 8806 | UpdateNodeOperands(TokenFactor.getNode(), OldChain, NewChain); |
| 8807 | return TokenFactor; |
| 8808 | } |
| 8809 | |
| 8810 | SDValue SelectionDAG::getSymbolFunctionGlobalAddress(SDValue Op, |
| 8811 | Function **OutFunction) { |
| 8812 | assert(isa<ExternalSymbolSDNode>(Op) && "Node should be an ExternalSymbol" ); |
| 8813 | |
| 8814 | auto *Symbol = cast<ExternalSymbolSDNode>(Op)->getSymbol(); |
| 8815 | auto *Module = MF->getFunction().getParent(); |
| 8816 | auto *Function = Module->getFunction(Symbol); |
| 8817 | |
| 8818 | if (OutFunction != nullptr) |
| 8819 | *OutFunction = Function; |
| 8820 | |
| 8821 | if (Function != nullptr) { |
| 8822 | auto PtrTy = TLI->getPointerTy(getDataLayout(), Function->getAddressSpace()); |
| 8823 | return getGlobalAddress(Function, SDLoc(Op), PtrTy); |
| 8824 | } |
| 8825 | |
| 8826 | std::string ErrorStr; |
| 8827 | raw_string_ostream ErrorFormatter(ErrorStr); |
| 8828 | |
| 8829 | ErrorFormatter << "Undefined external symbol " ; |
| 8830 | ErrorFormatter << '"' << Symbol << '"'; |
| 8831 | ErrorFormatter.flush(); |
| 8832 | |
| 8833 | report_fatal_error(ErrorStr); |
| 8834 | } |
| 8835 | |
| 8836 | //===----------------------------------------------------------------------===// |
| 8837 | // SDNode Class |
| 8838 | //===----------------------------------------------------------------------===// |
| 8839 | |
| 8840 | bool llvm::isNullConstant(SDValue V) { |
| 8841 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V); |
| 8842 | return Const != nullptr && Const->isNullValue(); |
| 8843 | } |
| 8844 | |
| 8845 | bool llvm::isNullFPConstant(SDValue V) { |
| 8846 | ConstantFPSDNode *Const = dyn_cast<ConstantFPSDNode>(V); |
| 8847 | return Const != nullptr && Const->isZero() && !Const->isNegative(); |
| 8848 | } |
| 8849 | |
| 8850 | bool llvm::isAllOnesConstant(SDValue V) { |
| 8851 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V); |
| 8852 | return Const != nullptr && Const->isAllOnesValue(); |
| 8853 | } |
| 8854 | |
| 8855 | bool llvm::isOneConstant(SDValue V) { |
| 8856 | ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V); |
| 8857 | return Const != nullptr && Const->isOne(); |
| 8858 | } |
| 8859 | |
| 8860 | SDValue llvm::peekThroughBitcasts(SDValue V) { |
| 8861 | while (V.getOpcode() == ISD::BITCAST) |
| 8862 | V = V.getOperand(0); |
| 8863 | return V; |
| 8864 | } |
| 8865 | |
| 8866 | SDValue llvm::peekThroughOneUseBitcasts(SDValue V) { |
| 8867 | while (V.getOpcode() == ISD::BITCAST && V.getOperand(0).hasOneUse()) |
| 8868 | V = V.getOperand(0); |
| 8869 | return V; |
| 8870 | } |
| 8871 | |
| 8872 | SDValue llvm::(SDValue V) { |
| 8873 | while (V.getOpcode() == ISD::EXTRACT_SUBVECTOR) |
| 8874 | V = V.getOperand(0); |
| 8875 | return V; |
| 8876 | } |
| 8877 | |
| 8878 | bool llvm::isBitwiseNot(SDValue V, bool AllowUndefs) { |
| 8879 | if (V.getOpcode() != ISD::XOR) |
| 8880 | return false; |
| 8881 | V = peekThroughBitcasts(V.getOperand(1)); |
| 8882 | unsigned NumBits = V.getScalarValueSizeInBits(); |
| 8883 | ConstantSDNode *C = |
| 8884 | isConstOrConstSplat(V, AllowUndefs, /*AllowTruncation*/ true); |
| 8885 | return C && (C->getAPIntValue().countTrailingOnes() >= NumBits); |
| 8886 | } |
| 8887 | |
| 8888 | ConstantSDNode *llvm::isConstOrConstSplat(SDValue N, bool AllowUndefs, |
| 8889 | bool AllowTruncation) { |
| 8890 | if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) |
| 8891 | return CN; |
| 8892 | |
| 8893 | if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) { |
| 8894 | BitVector UndefElements; |
| 8895 | ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements); |
| 8896 | |
| 8897 | // BuildVectors can truncate their operands. Ignore that case here unless |
| 8898 | // AllowTruncation is set. |
| 8899 | if (CN && (UndefElements.none() || AllowUndefs)) { |
| 8900 | EVT CVT = CN->getValueType(0); |
| 8901 | EVT NSVT = N.getValueType().getScalarType(); |
| 8902 | assert(CVT.bitsGE(NSVT) && "Illegal build vector element extension" ); |
| 8903 | if (AllowTruncation || (CVT == NSVT)) |
| 8904 | return CN; |
| 8905 | } |
| 8906 | } |
| 8907 | |
| 8908 | return nullptr; |
| 8909 | } |
| 8910 | |
| 8911 | ConstantSDNode *llvm::isConstOrConstSplat(SDValue N, const APInt &DemandedElts, |
| 8912 | bool AllowUndefs, |
| 8913 | bool AllowTruncation) { |
| 8914 | if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) |
| 8915 | return CN; |
| 8916 | |
| 8917 | if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) { |
| 8918 | BitVector UndefElements; |
| 8919 | ConstantSDNode *CN = BV->getConstantSplatNode(DemandedElts, &UndefElements); |
| 8920 | |
| 8921 | // BuildVectors can truncate their operands. Ignore that case here unless |
| 8922 | // AllowTruncation is set. |
| 8923 | if (CN && (UndefElements.none() || AllowUndefs)) { |
| 8924 | EVT CVT = CN->getValueType(0); |
| 8925 | EVT NSVT = N.getValueType().getScalarType(); |
| 8926 | assert(CVT.bitsGE(NSVT) && "Illegal build vector element extension" ); |
| 8927 | if (AllowTruncation || (CVT == NSVT)) |
| 8928 | return CN; |
| 8929 | } |
| 8930 | } |
| 8931 | |
| 8932 | return nullptr; |
| 8933 | } |
| 8934 | |
| 8935 | ConstantFPSDNode *llvm::isConstOrConstSplatFP(SDValue N, bool AllowUndefs) { |
| 8936 | if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N)) |
| 8937 | return CN; |
| 8938 | |
| 8939 | if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) { |
| 8940 | BitVector UndefElements; |
| 8941 | ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements); |
| 8942 | if (CN && (UndefElements.none() || AllowUndefs)) |
| 8943 | return CN; |
| 8944 | } |
| 8945 | |
| 8946 | return nullptr; |
| 8947 | } |
| 8948 | |
| 8949 | ConstantFPSDNode *llvm::isConstOrConstSplatFP(SDValue N, |
| 8950 | const APInt &DemandedElts, |
| 8951 | bool AllowUndefs) { |
| 8952 | if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N)) |
| 8953 | return CN; |
| 8954 | |
| 8955 | if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) { |
| 8956 | BitVector UndefElements; |
| 8957 | ConstantFPSDNode *CN = |
| 8958 | BV->getConstantFPSplatNode(DemandedElts, &UndefElements); |
| 8959 | if (CN && (UndefElements.none() || AllowUndefs)) |
| 8960 | return CN; |
| 8961 | } |
| 8962 | |
| 8963 | return nullptr; |
| 8964 | } |
| 8965 | |
| 8966 | bool llvm::isNullOrNullSplat(SDValue N, bool AllowUndefs) { |
| 8967 | // TODO: may want to use peekThroughBitcast() here. |
| 8968 | ConstantSDNode *C = isConstOrConstSplat(N, AllowUndefs); |
| 8969 | return C && C->isNullValue(); |
| 8970 | } |
| 8971 | |
| 8972 | bool llvm::isOneOrOneSplat(SDValue N) { |
| 8973 | // TODO: may want to use peekThroughBitcast() here. |
| 8974 | unsigned BitWidth = N.getScalarValueSizeInBits(); |
| 8975 | ConstantSDNode *C = isConstOrConstSplat(N); |
| 8976 | return C && C->isOne() && C->getValueSizeInBits(0) == BitWidth; |
| 8977 | } |
| 8978 | |
| 8979 | bool llvm::isAllOnesOrAllOnesSplat(SDValue N) { |
| 8980 | N = peekThroughBitcasts(N); |
| 8981 | unsigned BitWidth = N.getScalarValueSizeInBits(); |
| 8982 | ConstantSDNode *C = isConstOrConstSplat(N); |
| 8983 | return C && C->isAllOnesValue() && C->getValueSizeInBits(0) == BitWidth; |
| 8984 | } |
| 8985 | |
| 8986 | HandleSDNode::~HandleSDNode() { |
| 8987 | DropOperands(); |
| 8988 | } |
| 8989 | |
| 8990 | GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, unsigned Order, |
| 8991 | const DebugLoc &DL, |
| 8992 | const GlobalValue *GA, EVT VT, |
| 8993 | int64_t o, unsigned char TF) |
| 8994 | : SDNode(Opc, Order, DL, getSDVTList(VT)), Offset(o), TargetFlags(TF) { |
| 8995 | TheGlobal = GA; |
| 8996 | } |
| 8997 | |
| 8998 | AddrSpaceCastSDNode::AddrSpaceCastSDNode(unsigned Order, const DebugLoc &dl, |
| 8999 | EVT VT, unsigned SrcAS, |
| 9000 | unsigned DestAS) |
| 9001 | : SDNode(ISD::ADDRSPACECAST, Order, dl, getSDVTList(VT)), |
| 9002 | SrcAddrSpace(SrcAS), DestAddrSpace(DestAS) {} |
| 9003 | |
| 9004 | MemSDNode::MemSDNode(unsigned Opc, unsigned Order, const DebugLoc &dl, |
| 9005 | SDVTList VTs, EVT memvt, MachineMemOperand *mmo) |
| 9006 | : SDNode(Opc, Order, dl, VTs), MemoryVT(memvt), MMO(mmo) { |
| 9007 | MemSDNodeBits.IsVolatile = MMO->isVolatile(); |
| 9008 | MemSDNodeBits.IsNonTemporal = MMO->isNonTemporal(); |
| 9009 | MemSDNodeBits.IsDereferenceable = MMO->isDereferenceable(); |
| 9010 | MemSDNodeBits.IsInvariant = MMO->isInvariant(); |
| 9011 | |
| 9012 | // We check here that the size of the memory operand fits within the size of |
| 9013 | // the MMO. This is because the MMO might indicate only a possible address |
| 9014 | // range instead of specifying the affected memory addresses precisely. |
| 9015 | assert(memvt.getStoreSize() <= MMO->getSize() && "Size mismatch!" ); |
| 9016 | } |
| 9017 | |
| 9018 | /// Profile - Gather unique data for the node. |
| 9019 | /// |
| 9020 | void SDNode::Profile(FoldingSetNodeID &ID) const { |
| 9021 | AddNodeIDNode(ID, this); |
| 9022 | } |
| 9023 | |
| 9024 | namespace { |
| 9025 | |
| 9026 | struct EVTArray { |
| 9027 | std::vector<EVT> VTs; |
| 9028 | |
| 9029 | EVTArray() { |
| 9030 | VTs.reserve(MVT::LAST_VALUETYPE); |
| 9031 | for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i) |
| 9032 | VTs.push_back(MVT((MVT::SimpleValueType)i)); |
| 9033 | } |
| 9034 | }; |
| 9035 | |
| 9036 | } // end anonymous namespace |
| 9037 | |
| 9038 | static ManagedStatic<std::set<EVT, EVT::compareRawBits>> EVTs; |
| 9039 | static ManagedStatic<EVTArray> SimpleVTArray; |
| 9040 | static ManagedStatic<sys::SmartMutex<true>> VTMutex; |
| 9041 | |
| 9042 | /// getValueTypeList - Return a pointer to the specified value type. |
| 9043 | /// |
| 9044 | const EVT *SDNode::getValueTypeList(EVT VT) { |
| 9045 | if (VT.isExtended()) { |
| 9046 | sys::SmartScopedLock<true> Lock(*VTMutex); |
| 9047 | return &(*EVTs->insert(VT).first); |
| 9048 | } else { |
| 9049 | assert(VT.getSimpleVT() < MVT::LAST_VALUETYPE && |
| 9050 | "Value type out of range!" ); |
| 9051 | return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy]; |
| 9052 | } |
| 9053 | } |
| 9054 | |
| 9055 | /// hasNUsesOfValue - Return true if there are exactly NUSES uses of the |
| 9056 | /// indicated value. This method ignores uses of other values defined by this |
| 9057 | /// operation. |
| 9058 | bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const { |
| 9059 | assert(Value < getNumValues() && "Bad value!" ); |
| 9060 | |
| 9061 | // TODO: Only iterate over uses of a given value of the node |
| 9062 | for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) { |
| 9063 | if (UI.getUse().getResNo() == Value) { |
| 9064 | if (NUses == 0) |
| 9065 | return false; |
| 9066 | --NUses; |
| 9067 | } |
| 9068 | } |
| 9069 | |
| 9070 | // Found exactly the right number of uses? |
| 9071 | return NUses == 0; |
| 9072 | } |
| 9073 | |
| 9074 | /// hasAnyUseOfValue - Return true if there are any use of the indicated |
| 9075 | /// value. This method ignores uses of other values defined by this operation. |
| 9076 | bool SDNode::hasAnyUseOfValue(unsigned Value) const { |
| 9077 | assert(Value < getNumValues() && "Bad value!" ); |
| 9078 | |
| 9079 | for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) |
| 9080 | if (UI.getUse().getResNo() == Value) |
| 9081 | return true; |
| 9082 | |
| 9083 | return false; |
| 9084 | } |
| 9085 | |
| 9086 | /// isOnlyUserOf - Return true if this node is the only use of N. |
| 9087 | bool SDNode::isOnlyUserOf(const SDNode *N) const { |
| 9088 | bool Seen = false; |
| 9089 | for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) { |
| 9090 | SDNode *User = *I; |
| 9091 | if (User == this) |
| 9092 | Seen = true; |
| 9093 | else |
| 9094 | return false; |
| 9095 | } |
| 9096 | |
| 9097 | return Seen; |
| 9098 | } |
| 9099 | |
| 9100 | /// Return true if the only users of N are contained in Nodes. |
| 9101 | bool SDNode::areOnlyUsersOf(ArrayRef<const SDNode *> Nodes, const SDNode *N) { |
| 9102 | bool Seen = false; |
| 9103 | for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) { |
| 9104 | SDNode *User = *I; |
| 9105 | if (llvm::any_of(Nodes, |
| 9106 | [&User](const SDNode *Node) { return User == Node; })) |
| 9107 | Seen = true; |
| 9108 | else |
| 9109 | return false; |
| 9110 | } |
| 9111 | |
| 9112 | return Seen; |
| 9113 | } |
| 9114 | |
| 9115 | /// isOperand - Return true if this node is an operand of N. |
| 9116 | bool SDValue::isOperandOf(const SDNode *N) const { |
| 9117 | return any_of(N->op_values(), [this](SDValue Op) { return *this == Op; }); |
| 9118 | } |
| 9119 | |
| 9120 | bool SDNode::isOperandOf(const SDNode *N) const { |
| 9121 | return any_of(N->op_values(), |
| 9122 | [this](SDValue Op) { return this == Op.getNode(); }); |
| 9123 | } |
| 9124 | |
| 9125 | /// reachesChainWithoutSideEffects - Return true if this operand (which must |
| 9126 | /// be a chain) reaches the specified operand without crossing any |
| 9127 | /// side-effecting instructions on any chain path. In practice, this looks |
| 9128 | /// through token factors and non-volatile loads. In order to remain efficient, |
| 9129 | /// this only looks a couple of nodes in, it does not do an exhaustive search. |
| 9130 | /// |
| 9131 | /// Note that we only need to examine chains when we're searching for |
| 9132 | /// side-effects; SelectionDAG requires that all side-effects are represented |
| 9133 | /// by chains, even if another operand would force a specific ordering. This |
| 9134 | /// constraint is necessary to allow transformations like splitting loads. |
| 9135 | bool SDValue::reachesChainWithoutSideEffects(SDValue Dest, |
| 9136 | unsigned Depth) const { |
| 9137 | if (*this == Dest) return true; |
| 9138 | |
| 9139 | // Don't search too deeply, we just want to be able to see through |
| 9140 | // TokenFactor's etc. |
| 9141 | if (Depth == 0) return false; |
| 9142 | |
| 9143 | // If this is a token factor, all inputs to the TF happen in parallel. |
| 9144 | if (getOpcode() == ISD::TokenFactor) { |
| 9145 | // First, try a shallow search. |
| 9146 | if (is_contained((*this)->ops(), Dest)) { |
| 9147 | // We found the chain we want as an operand of this TokenFactor. |
| 9148 | // Essentially, we reach the chain without side-effects if we could |
| 9149 | // serialize the TokenFactor into a simple chain of operations with |
| 9150 | // Dest as the last operation. This is automatically true if the |
| 9151 | // chain has one use: there are no other ordering constraints. |
| 9152 | // If the chain has more than one use, we give up: some other |
| 9153 | // use of Dest might force a side-effect between Dest and the current |
| 9154 | // node. |
| 9155 | if (Dest.hasOneUse()) |
| 9156 | return true; |
| 9157 | } |
| 9158 | // Next, try a deep search: check whether every operand of the TokenFactor |
| 9159 | // reaches Dest. |
| 9160 | return llvm::all_of((*this)->ops(), [=](SDValue Op) { |
| 9161 | return Op.reachesChainWithoutSideEffects(Dest, Depth - 1); |
| 9162 | }); |
| 9163 | } |
| 9164 | |
| 9165 | // Loads don't have side effects, look through them. |
| 9166 | if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) { |
| 9167 | if (!Ld->isVolatile()) |
| 9168 | return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1); |
| 9169 | } |
| 9170 | return false; |
| 9171 | } |
| 9172 | |
| 9173 | bool SDNode::hasPredecessor(const SDNode *N) const { |
| 9174 | SmallPtrSet<const SDNode *, 32> Visited; |
| 9175 | SmallVector<const SDNode *, 16> Worklist; |
| 9176 | Worklist.push_back(this); |
| 9177 | return hasPredecessorHelper(N, Visited, Worklist); |
| 9178 | } |
| 9179 | |
| 9180 | void SDNode::intersectFlagsWith(const SDNodeFlags Flags) { |
| 9181 | this->Flags.intersectWith(Flags); |
| 9182 | } |
| 9183 | |
| 9184 | SDValue |
| 9185 | SelectionDAG::matchBinOpReduction(SDNode *, ISD::NodeType &BinOp, |
| 9186 | ArrayRef<ISD::NodeType> CandidateBinOps) { |
| 9187 | // The pattern must end in an extract from index 0. |
| 9188 | if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT || |
| 9189 | !isNullConstant(Extract->getOperand(1))) |
| 9190 | return SDValue(); |
| 9191 | |
| 9192 | SDValue Op = Extract->getOperand(0); |
| 9193 | unsigned Stages = Log2_32(Op.getValueType().getVectorNumElements()); |
| 9194 | |
| 9195 | // Match against one of the candidate binary ops. |
| 9196 | if (llvm::none_of(CandidateBinOps, [Op](ISD::NodeType BinOp) { |
| 9197 | return Op.getOpcode() == unsigned(BinOp); |
| 9198 | })) |
| 9199 | return SDValue(); |
| 9200 | |
| 9201 | // At each stage, we're looking for something that looks like: |
| 9202 | // %s = shufflevector <8 x i32> %op, <8 x i32> undef, |
| 9203 | // <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, |
| 9204 | // i32 undef, i32 undef, i32 undef, i32 undef> |
| 9205 | // %a = binop <8 x i32> %op, %s |
| 9206 | // Where the mask changes according to the stage. E.g. for a 3-stage pyramid, |
| 9207 | // we expect something like: |
| 9208 | // <4,5,6,7,u,u,u,u> |
| 9209 | // <2,3,u,u,u,u,u,u> |
| 9210 | // <1,u,u,u,u,u,u,u> |
| 9211 | unsigned CandidateBinOp = Op.getOpcode(); |
| 9212 | for (unsigned i = 0; i < Stages; ++i) { |
| 9213 | if (Op.getOpcode() != CandidateBinOp) |
| 9214 | return SDValue(); |
| 9215 | |
| 9216 | SDValue Op0 = Op.getOperand(0); |
| 9217 | SDValue Op1 = Op.getOperand(1); |
| 9218 | |
| 9219 | ShuffleVectorSDNode *Shuffle = dyn_cast<ShuffleVectorSDNode>(Op0); |
| 9220 | if (Shuffle) { |
| 9221 | Op = Op1; |
| 9222 | } else { |
| 9223 | Shuffle = dyn_cast<ShuffleVectorSDNode>(Op1); |
| 9224 | Op = Op0; |
| 9225 | } |
| 9226 | |
| 9227 | // The first operand of the shuffle should be the same as the other operand |
| 9228 | // of the binop. |
| 9229 | if (!Shuffle || Shuffle->getOperand(0) != Op) |
| 9230 | return SDValue(); |
| 9231 | |
| 9232 | // Verify the shuffle has the expected (at this stage of the pyramid) mask. |
| 9233 | for (int Index = 0, MaskEnd = 1 << i; Index < MaskEnd; ++Index) |
| 9234 | if (Shuffle->getMaskElt(Index) != MaskEnd + Index) |
| 9235 | return SDValue(); |
| 9236 | } |
| 9237 | |
| 9238 | BinOp = (ISD::NodeType)CandidateBinOp; |
| 9239 | return Op; |
| 9240 | } |
| 9241 | |
| 9242 | SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) { |
| 9243 | assert(N->getNumValues() == 1 && |
| 9244 | "Can't unroll a vector with multiple results!" ); |
| 9245 | |
| 9246 | EVT VT = N->getValueType(0); |
| 9247 | unsigned NE = VT.getVectorNumElements(); |
| 9248 | EVT EltVT = VT.getVectorElementType(); |
| 9249 | SDLoc dl(N); |
| 9250 | |
| 9251 | SmallVector<SDValue, 8> Scalars; |
| 9252 | SmallVector<SDValue, 4> Operands(N->getNumOperands()); |
| 9253 | |
| 9254 | // If ResNE is 0, fully unroll the vector op. |
| 9255 | if (ResNE == 0) |
| 9256 | ResNE = NE; |
| 9257 | else if (NE > ResNE) |
| 9258 | NE = ResNE; |
| 9259 | |
| 9260 | unsigned i; |
| 9261 | for (i= 0; i != NE; ++i) { |
| 9262 | for (unsigned j = 0, e = N->getNumOperands(); j != e; ++j) { |
| 9263 | SDValue Operand = N->getOperand(j); |
| 9264 | EVT OperandVT = Operand.getValueType(); |
| 9265 | if (OperandVT.isVector()) { |
| 9266 | // A vector operand; extract a single element. |
| 9267 | EVT OperandEltVT = OperandVT.getVectorElementType(); |
| 9268 | Operands[j] = |
| 9269 | getNode(ISD::EXTRACT_VECTOR_ELT, dl, OperandEltVT, Operand, |
| 9270 | getConstant(i, dl, TLI->getVectorIdxTy(getDataLayout()))); |
| 9271 | } else { |
| 9272 | // A scalar operand; just use it as is. |
| 9273 | Operands[j] = Operand; |
| 9274 | } |
| 9275 | } |
| 9276 | |
| 9277 | switch (N->getOpcode()) { |
| 9278 | default: { |
| 9279 | Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands, |
| 9280 | N->getFlags())); |
| 9281 | break; |
| 9282 | } |
| 9283 | case ISD::VSELECT: |
| 9284 | Scalars.push_back(getNode(ISD::SELECT, dl, EltVT, Operands)); |
| 9285 | break; |
| 9286 | case ISD::SHL: |
| 9287 | case ISD::SRA: |
| 9288 | case ISD::SRL: |
| 9289 | case ISD::ROTL: |
| 9290 | case ISD::ROTR: |
| 9291 | Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0], |
| 9292 | getShiftAmountOperand(Operands[0].getValueType(), |
| 9293 | Operands[1]))); |
| 9294 | break; |
| 9295 | case ISD::SIGN_EXTEND_INREG: |
| 9296 | case ISD::FP_ROUND_INREG: { |
| 9297 | EVT ExtVT = cast<VTSDNode>(Operands[1])->getVT().getVectorElementType(); |
| 9298 | Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, |
| 9299 | Operands[0], |
| 9300 | getValueType(ExtVT))); |
| 9301 | } |
| 9302 | } |
| 9303 | } |
| 9304 | |
| 9305 | for (; i < ResNE; ++i) |
| 9306 | Scalars.push_back(getUNDEF(EltVT)); |
| 9307 | |
| 9308 | EVT VecVT = EVT::getVectorVT(*getContext(), EltVT, ResNE); |
| 9309 | return getBuildVector(VecVT, dl, Scalars); |
| 9310 | } |
| 9311 | |
| 9312 | std::pair<SDValue, SDValue> SelectionDAG::UnrollVectorOverflowOp( |
| 9313 | SDNode *N, unsigned ResNE) { |
| 9314 | unsigned Opcode = N->getOpcode(); |
| 9315 | assert((Opcode == ISD::UADDO || Opcode == ISD::SADDO || |
| 9316 | Opcode == ISD::USUBO || Opcode == ISD::SSUBO || |
| 9317 | Opcode == ISD::UMULO || Opcode == ISD::SMULO) && |
| 9318 | "Expected an overflow opcode" ); |
| 9319 | |
| 9320 | EVT ResVT = N->getValueType(0); |
| 9321 | EVT OvVT = N->getValueType(1); |
| 9322 | EVT ResEltVT = ResVT.getVectorElementType(); |
| 9323 | EVT OvEltVT = OvVT.getVectorElementType(); |
| 9324 | SDLoc dl(N); |
| 9325 | |
| 9326 | // If ResNE is 0, fully unroll the vector op. |
| 9327 | unsigned NE = ResVT.getVectorNumElements(); |
| 9328 | if (ResNE == 0) |
| 9329 | ResNE = NE; |
| 9330 | else if (NE > ResNE) |
| 9331 | NE = ResNE; |
| 9332 | |
| 9333 | SmallVector<SDValue, 8> LHSScalars; |
| 9334 | SmallVector<SDValue, 8> RHSScalars; |
| 9335 | ExtractVectorElements(N->getOperand(0), LHSScalars, 0, NE); |
| 9336 | ExtractVectorElements(N->getOperand(1), RHSScalars, 0, NE); |
| 9337 | |
| 9338 | EVT SVT = TLI->getSetCCResultType(getDataLayout(), *getContext(), ResEltVT); |
| 9339 | SDVTList VTs = getVTList(ResEltVT, SVT); |
| 9340 | SmallVector<SDValue, 8> ResScalars; |
| 9341 | SmallVector<SDValue, 8> OvScalars; |
| 9342 | for (unsigned i = 0; i < NE; ++i) { |
| 9343 | SDValue Res = getNode(Opcode, dl, VTs, LHSScalars[i], RHSScalars[i]); |
| 9344 | SDValue Ov = |
| 9345 | getSelect(dl, OvEltVT, Res.getValue(1), |
| 9346 | getBoolConstant(true, dl, OvEltVT, ResVT), |
| 9347 | getConstant(0, dl, OvEltVT)); |
| 9348 | |
| 9349 | ResScalars.push_back(Res); |
| 9350 | OvScalars.push_back(Ov); |
| 9351 | } |
| 9352 | |
| 9353 | ResScalars.append(ResNE - NE, getUNDEF(ResEltVT)); |
| 9354 | OvScalars.append(ResNE - NE, getUNDEF(OvEltVT)); |
| 9355 | |
| 9356 | EVT NewResVT = EVT::getVectorVT(*getContext(), ResEltVT, ResNE); |
| 9357 | EVT NewOvVT = EVT::getVectorVT(*getContext(), OvEltVT, ResNE); |
| 9358 | return std::make_pair(getBuildVector(NewResVT, dl, ResScalars), |
| 9359 | getBuildVector(NewOvVT, dl, OvScalars)); |
| 9360 | } |
| 9361 | |
| 9362 | bool SelectionDAG::areNonVolatileConsecutiveLoads(LoadSDNode *LD, |
| 9363 | LoadSDNode *Base, |
| 9364 | unsigned Bytes, |
| 9365 | int Dist) const { |
| 9366 | if (LD->isVolatile() || Base->isVolatile()) |
| 9367 | return false; |
| 9368 | if (LD->isIndexed() || Base->isIndexed()) |
| 9369 | return false; |
| 9370 | if (LD->getChain() != Base->getChain()) |
| 9371 | return false; |
| 9372 | EVT VT = LD->getValueType(0); |
| 9373 | if (VT.getSizeInBits() / 8 != Bytes) |
| 9374 | return false; |
| 9375 | |
| 9376 | auto BaseLocDecomp = BaseIndexOffset::match(Base, *this); |
| 9377 | auto LocDecomp = BaseIndexOffset::match(LD, *this); |
| 9378 | |
| 9379 | int64_t Offset = 0; |
| 9380 | if (BaseLocDecomp.equalBaseIndex(LocDecomp, *this, Offset)) |
| 9381 | return (Dist * Bytes == Offset); |
| 9382 | return false; |
| 9383 | } |
| 9384 | |
| 9385 | /// InferPtrAlignment - Infer alignment of a load / store address. Return 0 if |
| 9386 | /// it cannot be inferred. |
| 9387 | unsigned SelectionDAG::InferPtrAlignment(SDValue Ptr) const { |
| 9388 | // If this is a GlobalAddress + cst, return the alignment. |
| 9389 | const GlobalValue *GV; |
| 9390 | int64_t GVOffset = 0; |
| 9391 | if (TLI->isGAPlusOffset(Ptr.getNode(), GV, GVOffset)) { |
| 9392 | unsigned IdxWidth = getDataLayout().getIndexTypeSizeInBits(GV->getType()); |
| 9393 | KnownBits Known(IdxWidth); |
| 9394 | llvm::computeKnownBits(GV, Known, getDataLayout()); |
| 9395 | unsigned AlignBits = Known.countMinTrailingZeros(); |
| 9396 | unsigned Align = AlignBits ? 1 << std::min(31U, AlignBits) : 0; |
| 9397 | if (Align) |
| 9398 | return MinAlign(Align, GVOffset); |
| 9399 | } |
| 9400 | |
| 9401 | // If this is a direct reference to a stack slot, use information about the |
| 9402 | // stack slot's alignment. |
| 9403 | int FrameIdx = INT_MIN; |
| 9404 | int64_t FrameOffset = 0; |
| 9405 | if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) { |
| 9406 | FrameIdx = FI->getIndex(); |
| 9407 | } else if (isBaseWithConstantOffset(Ptr) && |
| 9408 | isa<FrameIndexSDNode>(Ptr.getOperand(0))) { |
| 9409 | // Handle FI+Cst |
| 9410 | FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); |
| 9411 | FrameOffset = Ptr.getConstantOperandVal(1); |
| 9412 | } |
| 9413 | |
| 9414 | if (FrameIdx != INT_MIN) { |
| 9415 | const MachineFrameInfo &MFI = getMachineFunction().getFrameInfo(); |
| 9416 | unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx), |
| 9417 | FrameOffset); |
| 9418 | return FIInfoAlign; |
| 9419 | } |
| 9420 | |
| 9421 | return 0; |
| 9422 | } |
| 9423 | |
| 9424 | /// GetSplitDestVTs - Compute the VTs needed for the low/hi parts of a type |
| 9425 | /// which is split (or expanded) into two not necessarily identical pieces. |
| 9426 | std::pair<EVT, EVT> SelectionDAG::GetSplitDestVTs(const EVT &VT) const { |
| 9427 | // Currently all types are split in half. |
| 9428 | EVT LoVT, HiVT; |
| 9429 | if (!VT.isVector()) |
| 9430 | LoVT = HiVT = TLI->getTypeToTransformTo(*getContext(), VT); |
| 9431 | else |
| 9432 | LoVT = HiVT = VT.getHalfNumVectorElementsVT(*getContext()); |
| 9433 | |
| 9434 | return std::make_pair(LoVT, HiVT); |
| 9435 | } |
| 9436 | |
| 9437 | /// SplitVector - Split the vector with EXTRACT_SUBVECTOR and return the |
| 9438 | /// low/high part. |
| 9439 | std::pair<SDValue, SDValue> |
| 9440 | SelectionDAG::SplitVector(const SDValue &N, const SDLoc &DL, const EVT &LoVT, |
| 9441 | const EVT &HiVT) { |
| 9442 | assert(LoVT.getVectorNumElements() + HiVT.getVectorNumElements() <= |
| 9443 | N.getValueType().getVectorNumElements() && |
| 9444 | "More vector elements requested than available!" ); |
| 9445 | SDValue Lo, Hi; |
| 9446 | Lo = getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, N, |
| 9447 | getConstant(0, DL, TLI->getVectorIdxTy(getDataLayout()))); |
| 9448 | Hi = getNode(ISD::EXTRACT_SUBVECTOR, DL, HiVT, N, |
| 9449 | getConstant(LoVT.getVectorNumElements(), DL, |
| 9450 | TLI->getVectorIdxTy(getDataLayout()))); |
| 9451 | return std::make_pair(Lo, Hi); |
| 9452 | } |
| 9453 | |
| 9454 | /// Widen the vector up to the next power of two using INSERT_SUBVECTOR. |
| 9455 | SDValue SelectionDAG::WidenVector(const SDValue &N, const SDLoc &DL) { |
| 9456 | EVT VT = N.getValueType(); |
| 9457 | EVT WideVT = EVT::getVectorVT(*getContext(), VT.getVectorElementType(), |
| 9458 | NextPowerOf2(VT.getVectorNumElements())); |
| 9459 | return getNode(ISD::INSERT_SUBVECTOR, DL, WideVT, getUNDEF(WideVT), N, |
| 9460 | getConstant(0, DL, TLI->getVectorIdxTy(getDataLayout()))); |
| 9461 | } |
| 9462 | |
| 9463 | void SelectionDAG::(SDValue Op, |
| 9464 | SmallVectorImpl<SDValue> &Args, |
| 9465 | unsigned Start, unsigned Count) { |
| 9466 | EVT VT = Op.getValueType(); |
| 9467 | if (Count == 0) |
| 9468 | Count = VT.getVectorNumElements(); |
| 9469 | |
| 9470 | EVT EltVT = VT.getVectorElementType(); |
| 9471 | EVT IdxTy = TLI->getVectorIdxTy(getDataLayout()); |
| 9472 | SDLoc SL(Op); |
| 9473 | for (unsigned i = Start, e = Start + Count; i != e; ++i) { |
| 9474 | Args.push_back(getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT, |
| 9475 | Op, getConstant(i, SL, IdxTy))); |
| 9476 | } |
| 9477 | } |
| 9478 | |
| 9479 | // getAddressSpace - Return the address space this GlobalAddress belongs to. |
| 9480 | unsigned GlobalAddressSDNode::getAddressSpace() const { |
| 9481 | return getGlobal()->getType()->getAddressSpace(); |
| 9482 | } |
| 9483 | |
| 9484 | Type *ConstantPoolSDNode::getType() const { |
| 9485 | if (isMachineConstantPoolEntry()) |
| 9486 | return Val.MachineCPVal->getType(); |
| 9487 | return Val.ConstVal->getType(); |
| 9488 | } |
| 9489 | |
| 9490 | bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue, APInt &SplatUndef, |
| 9491 | unsigned &SplatBitSize, |
| 9492 | bool &HasAnyUndefs, |
| 9493 | unsigned MinSplatBits, |
| 9494 | bool IsBigEndian) const { |
| 9495 | EVT VT = getValueType(0); |
| 9496 | assert(VT.isVector() && "Expected a vector type" ); |
| 9497 | unsigned VecWidth = VT.getSizeInBits(); |
| 9498 | if (MinSplatBits > VecWidth) |
| 9499 | return false; |
| 9500 | |
| 9501 | // FIXME: The widths are based on this node's type, but build vectors can |
| 9502 | // truncate their operands. |
| 9503 | SplatValue = APInt(VecWidth, 0); |
| 9504 | SplatUndef = APInt(VecWidth, 0); |
| 9505 | |
| 9506 | // Get the bits. Bits with undefined values (when the corresponding element |
| 9507 | // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared |
| 9508 | // in SplatValue. If any of the values are not constant, give up and return |
| 9509 | // false. |
| 9510 | unsigned int NumOps = getNumOperands(); |
| 9511 | assert(NumOps > 0 && "isConstantSplat has 0-size build vector" ); |
| 9512 | unsigned EltWidth = VT.getScalarSizeInBits(); |
| 9513 | |
| 9514 | for (unsigned j = 0; j < NumOps; ++j) { |
| 9515 | unsigned i = IsBigEndian ? NumOps - 1 - j : j; |
| 9516 | SDValue OpVal = getOperand(i); |
| 9517 | unsigned BitPos = j * EltWidth; |
| 9518 | |
| 9519 | if (OpVal.isUndef()) |
| 9520 | SplatUndef.setBits(BitPos, BitPos + EltWidth); |
| 9521 | else if (auto *CN = dyn_cast<ConstantSDNode>(OpVal)) |
| 9522 | SplatValue.insertBits(CN->getAPIntValue().zextOrTrunc(EltWidth), BitPos); |
| 9523 | else if (auto *CN = dyn_cast<ConstantFPSDNode>(OpVal)) |
| 9524 | SplatValue.insertBits(CN->getValueAPF().bitcastToAPInt(), BitPos); |
| 9525 | else |
| 9526 | return false; |
| 9527 | } |
| 9528 | |
| 9529 | // The build_vector is all constants or undefs. Find the smallest element |
| 9530 | // size that splats the vector. |
| 9531 | HasAnyUndefs = (SplatUndef != 0); |
| 9532 | |
| 9533 | // FIXME: This does not work for vectors with elements less than 8 bits. |
| 9534 | while (VecWidth > 8) { |
| 9535 | unsigned HalfSize = VecWidth / 2; |
| 9536 | APInt HighValue = SplatValue.lshr(HalfSize).trunc(HalfSize); |
| 9537 | APInt LowValue = SplatValue.trunc(HalfSize); |
| 9538 | APInt HighUndef = SplatUndef.lshr(HalfSize).trunc(HalfSize); |
| 9539 | APInt LowUndef = SplatUndef.trunc(HalfSize); |
| 9540 | |
| 9541 | // If the two halves do not match (ignoring undef bits), stop here. |
| 9542 | if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) || |
| 9543 | MinSplatBits > HalfSize) |
| 9544 | break; |
| 9545 | |
| 9546 | SplatValue = HighValue | LowValue; |
| 9547 | SplatUndef = HighUndef & LowUndef; |
| 9548 | |
| 9549 | VecWidth = HalfSize; |
| 9550 | } |
| 9551 | |
| 9552 | SplatBitSize = VecWidth; |
| 9553 | return true; |
| 9554 | } |
| 9555 | |
| 9556 | SDValue BuildVectorSDNode::getSplatValue(const APInt &DemandedElts, |
| 9557 | BitVector *UndefElements) const { |
| 9558 | if (UndefElements) { |
| 9559 | UndefElements->clear(); |
| 9560 | UndefElements->resize(getNumOperands()); |
| 9561 | } |
| 9562 | assert(getNumOperands() == DemandedElts.getBitWidth() && |
| 9563 | "Unexpected vector size" ); |
| 9564 | if (!DemandedElts) |
| 9565 | return SDValue(); |
| 9566 | SDValue Splatted; |
| 9567 | for (unsigned i = 0, e = getNumOperands(); i != e; ++i) { |
| 9568 | if (!DemandedElts[i]) |
| 9569 | continue; |
| 9570 | SDValue Op = getOperand(i); |
| 9571 | if (Op.isUndef()) { |
| 9572 | if (UndefElements) |
| 9573 | (*UndefElements)[i] = true; |
| 9574 | } else if (!Splatted) { |
| 9575 | Splatted = Op; |
| 9576 | } else if (Splatted != Op) { |
| 9577 | return SDValue(); |
| 9578 | } |
| 9579 | } |
| 9580 | |
| 9581 | if (!Splatted) { |
| 9582 | unsigned FirstDemandedIdx = DemandedElts.countTrailingZeros(); |
| 9583 | assert(getOperand(FirstDemandedIdx).isUndef() && |
| 9584 | "Can only have a splat without a constant for all undefs." ); |
| 9585 | return getOperand(FirstDemandedIdx); |
| 9586 | } |
| 9587 | |
| 9588 | return Splatted; |
| 9589 | } |
| 9590 | |
| 9591 | SDValue BuildVectorSDNode::getSplatValue(BitVector *UndefElements) const { |
| 9592 | APInt DemandedElts = APInt::getAllOnesValue(getNumOperands()); |
| 9593 | return getSplatValue(DemandedElts, UndefElements); |
| 9594 | } |
| 9595 | |
| 9596 | ConstantSDNode * |
| 9597 | BuildVectorSDNode::getConstantSplatNode(const APInt &DemandedElts, |
| 9598 | BitVector *UndefElements) const { |
| 9599 | return dyn_cast_or_null<ConstantSDNode>( |
| 9600 | getSplatValue(DemandedElts, UndefElements)); |
| 9601 | } |
| 9602 | |
| 9603 | ConstantSDNode * |
| 9604 | BuildVectorSDNode::getConstantSplatNode(BitVector *UndefElements) const { |
| 9605 | return dyn_cast_or_null<ConstantSDNode>(getSplatValue(UndefElements)); |
| 9606 | } |
| 9607 | |
| 9608 | ConstantFPSDNode * |
| 9609 | BuildVectorSDNode::getConstantFPSplatNode(const APInt &DemandedElts, |
| 9610 | BitVector *UndefElements) const { |
| 9611 | return dyn_cast_or_null<ConstantFPSDNode>( |
| 9612 | getSplatValue(DemandedElts, UndefElements)); |
| 9613 | } |
| 9614 | |
| 9615 | ConstantFPSDNode * |
| 9616 | BuildVectorSDNode::getConstantFPSplatNode(BitVector *UndefElements) const { |
| 9617 | return dyn_cast_or_null<ConstantFPSDNode>(getSplatValue(UndefElements)); |
| 9618 | } |
| 9619 | |
| 9620 | int32_t |
| 9621 | BuildVectorSDNode::getConstantFPSplatPow2ToLog2Int(BitVector *UndefElements, |
| 9622 | uint32_t BitWidth) const { |
| 9623 | if (ConstantFPSDNode *CN = |
| 9624 | dyn_cast_or_null<ConstantFPSDNode>(getSplatValue(UndefElements))) { |
| 9625 | bool IsExact; |
| 9626 | APSInt IntVal(BitWidth); |
| 9627 | const APFloat &APF = CN->getValueAPF(); |
| 9628 | if (APF.convertToInteger(IntVal, APFloat::rmTowardZero, &IsExact) != |
| 9629 | APFloat::opOK || |
| 9630 | !IsExact) |
| 9631 | return -1; |
| 9632 | |
| 9633 | return IntVal.exactLogBase2(); |
| 9634 | } |
| 9635 | return -1; |
| 9636 | } |
| 9637 | |
| 9638 | bool BuildVectorSDNode::isConstant() const { |
| 9639 | for (const SDValue &Op : op_values()) { |
| 9640 | unsigned Opc = Op.getOpcode(); |
| 9641 | if (Opc != ISD::UNDEF && Opc != ISD::Constant && Opc != ISD::ConstantFP) |
| 9642 | return false; |
| 9643 | } |
| 9644 | return true; |
| 9645 | } |
| 9646 | |
| 9647 | bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) { |
| 9648 | // Find the first non-undef value in the shuffle mask. |
| 9649 | unsigned i, e; |
| 9650 | for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i) |
| 9651 | /* search */; |
| 9652 | |
| 9653 | // If all elements are undefined, this shuffle can be considered a splat |
| 9654 | // (although it should eventually get simplified away completely). |
| 9655 | if (i == e) |
| 9656 | return true; |
| 9657 | |
| 9658 | // Make sure all remaining elements are either undef or the same as the first |
| 9659 | // non-undef value. |
| 9660 | for (int Idx = Mask[i]; i != e; ++i) |
| 9661 | if (Mask[i] >= 0 && Mask[i] != Idx) |
| 9662 | return false; |
| 9663 | return true; |
| 9664 | } |
| 9665 | |
| 9666 | // Returns the SDNode if it is a constant integer BuildVector |
| 9667 | // or constant integer. |
| 9668 | SDNode *SelectionDAG::isConstantIntBuildVectorOrConstantInt(SDValue N) { |
| 9669 | if (isa<ConstantSDNode>(N)) |
| 9670 | return N.getNode(); |
| 9671 | if (ISD::isBuildVectorOfConstantSDNodes(N.getNode())) |
| 9672 | return N.getNode(); |
| 9673 | // Treat a GlobalAddress supporting constant offset folding as a |
| 9674 | // constant integer. |
| 9675 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N)) |
| 9676 | if (GA->getOpcode() == ISD::GlobalAddress && |
| 9677 | TLI->isOffsetFoldingLegal(GA)) |
| 9678 | return GA; |
| 9679 | return nullptr; |
| 9680 | } |
| 9681 | |
| 9682 | SDNode *SelectionDAG::isConstantFPBuildVectorOrConstantFP(SDValue N) { |
| 9683 | if (isa<ConstantFPSDNode>(N)) |
| 9684 | return N.getNode(); |
| 9685 | |
| 9686 | if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode())) |
| 9687 | return N.getNode(); |
| 9688 | |
| 9689 | return nullptr; |
| 9690 | } |
| 9691 | |
| 9692 | void SelectionDAG::createOperands(SDNode *Node, ArrayRef<SDValue> Vals) { |
| 9693 | assert(!Node->OperandList && "Node already has operands" ); |
| 9694 | assert(SDNode::getMaxNumOperands() >= Vals.size() && |
| 9695 | "too many operands to fit into SDNode" ); |
| 9696 | SDUse *Ops = OperandRecycler.allocate( |
| 9697 | ArrayRecycler<SDUse>::Capacity::get(Vals.size()), OperandAllocator); |
| 9698 | |
| 9699 | bool IsDivergent = false; |
| 9700 | for (unsigned I = 0; I != Vals.size(); ++I) { |
| 9701 | Ops[I].setUser(Node); |
| 9702 | Ops[I].setInitial(Vals[I]); |
| 9703 | if (Ops[I].Val.getValueType() != MVT::Other) // Skip Chain. It does not carry divergence. |
| 9704 | IsDivergent = IsDivergent || Ops[I].getNode()->isDivergent(); |
| 9705 | } |
| 9706 | Node->NumOperands = Vals.size(); |
| 9707 | Node->OperandList = Ops; |
| 9708 | IsDivergent |= TLI->isSDNodeSourceOfDivergence(Node, FLI, DA); |
| 9709 | if (!TLI->isSDNodeAlwaysUniform(Node)) |
| 9710 | Node->SDNodeBits.IsDivergent = IsDivergent; |
| 9711 | checkForCycles(Node); |
| 9712 | } |
| 9713 | |
| 9714 | SDValue SelectionDAG::getTokenFactor(const SDLoc &DL, |
| 9715 | SmallVectorImpl<SDValue> &Vals) { |
| 9716 | size_t Limit = SDNode::getMaxNumOperands(); |
| 9717 | while (Vals.size() > Limit) { |
| 9718 | unsigned SliceIdx = Vals.size() - Limit; |
| 9719 | auto = ArrayRef<SDValue>(Vals).slice(SliceIdx, Limit); |
| 9720 | SDValue NewTF = getNode(ISD::TokenFactor, DL, MVT::Other, ExtractedTFs); |
| 9721 | Vals.erase(Vals.begin() + SliceIdx, Vals.end()); |
| 9722 | Vals.emplace_back(NewTF); |
| 9723 | } |
| 9724 | return getNode(ISD::TokenFactor, DL, MVT::Other, Vals); |
| 9725 | } |
| 9726 | |
| 9727 | #ifndef NDEBUG |
| 9728 | static void checkForCyclesHelper(const SDNode *N, |
| 9729 | SmallPtrSetImpl<const SDNode*> &Visited, |
| 9730 | SmallPtrSetImpl<const SDNode*> &Checked, |
| 9731 | const llvm::SelectionDAG *DAG) { |
| 9732 | // If this node has already been checked, don't check it again. |
| 9733 | if (Checked.count(N)) |
| 9734 | return; |
| 9735 | |
| 9736 | // If a node has already been visited on this depth-first walk, reject it as |
| 9737 | // a cycle. |
| 9738 | if (!Visited.insert(N).second) { |
| 9739 | errs() << "Detected cycle in SelectionDAG\n" ; |
| 9740 | dbgs() << "Offending node:\n" ; |
| 9741 | N->dumprFull(DAG); dbgs() << "\n" ; |
| 9742 | abort(); |
| 9743 | } |
| 9744 | |
| 9745 | for (const SDValue &Op : N->op_values()) |
| 9746 | checkForCyclesHelper(Op.getNode(), Visited, Checked, DAG); |
| 9747 | |
| 9748 | Checked.insert(N); |
| 9749 | Visited.erase(N); |
| 9750 | } |
| 9751 | #endif |
| 9752 | |
| 9753 | void llvm::checkForCycles(const llvm::SDNode *N, |
| 9754 | const llvm::SelectionDAG *DAG, |
| 9755 | bool force) { |
| 9756 | #ifndef NDEBUG |
| 9757 | bool check = force; |
| 9758 | #ifdef EXPENSIVE_CHECKS |
| 9759 | check = true; |
| 9760 | #endif // EXPENSIVE_CHECKS |
| 9761 | if (check) { |
| 9762 | assert(N && "Checking nonexistent SDNode" ); |
| 9763 | SmallPtrSet<const SDNode*, 32> visited; |
| 9764 | SmallPtrSet<const SDNode*, 32> checked; |
| 9765 | checkForCyclesHelper(N, visited, checked, DAG); |
| 9766 | } |
| 9767 | #endif // !NDEBUG |
| 9768 | } |
| 9769 | |
| 9770 | void llvm::checkForCycles(const llvm::SelectionDAG *DAG, bool force) { |
| 9771 | checkForCycles(DAG->getRoot().getNode(), DAG, force); |
| 9772 | } |
| 9773 | |